Method for screening chemical hepatotoxicity based on high content cell imaging and convolutional neural network

Abstract

The invention provides a method for screening hepatotoxicity of chemicals based on high-content cell imaging and a convolutional neural network. The method comprises the following steps: constructing a high-content cell image data set of the chemicals; the method comprises the following steps: randomly selecting a plurality of images, training a cell nucleus and cell segmentation model suitable for a chemical high-content cell image data set by using Cellpose2.0, carrying out batch processing on the images by using the trained segmentation model, and storing a processed Mask image; based on the Mask image, phenotypic parameters are extracted and preprocessed; constructing a classification model based on the convolutional neural network by taking the preprocessed phenotypic parameters as input and taking existence of hepatotoxicity as an end point of the model; and carrying out hepatotoxicity screening by adopting the constructed classification model to obtain a prediction effect. According to the method disclosed by the invention, high-throughput screening of the hepatotoxicity of the chemicals is effectively realized, the established screening model has good fitting ability, robustness and predictive ability, an effective tool is provided for risk assessment and management of the chemicals, and the method has important application value.

Description

Technical Field

[0001] The present invention relates to the technical field of screening for risk assessment of chemical hepatotoxicity, and in particular to a method for screening chemical hepatotoxicity based on high-content cell imaging and convolutional neural networks. Background Art

[0002] The number of chemicals with hepatotoxicity data available is much lower than the current number of chemicals, and there is a huge gap in hepatotoxicity data for chemicals.

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

[0004] As a powerful deep learning model, convolutional neural network (CNN) can more deeply explore the complex patterns and structures in the data through automatic feature extraction and hierarchical feature representation, so as to process high-dimensional data more accurately and efficiently. High-content imaging technology captures many different microscopic descriptors by imaging stained cells, which cover the phenotypic characteristics of cells and their organelles. These phenotypic characteristics not only provide more information than traditional measurements, but also show the response of cells to chemicals from a broader perspective. Using phenotypic parameters as input to construct a convolutional neural network screening model for chemical hepatotoxicity is expected to fill the gap in hepatotoxicity data for many chemicals on the market.

[0005] At present, there are literatures that use high-content phenotypic parameters to build machine learning screening models in toxicological assessment. The literature "J Appl Toxicol. 2014, 34 (3), 281-8." uses the phenotypic parameters after cell imaging as input to build a random forest classification model for screening drug hepatotoxicity; the literature "Chem. Res. Toxicol. 2020, 33, 2261-2275." builds a drug hepatotoxicity prediction model based on high-content phenotyping and random forest; the literature "Chem Res Toxicol. 2023, 18, 36 (9), 1456–1470." uses high-content phenotyping 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 use 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 liver cancer cells). Summary of the invention

[0006] According to the technical problems raised above, a method for screening chemical hepatotoxicity based on high-content cell imaging and convolutional neural network is provided. The present invention selects high-content images of HepG2 cells, extracts phenotypic parameters from them, applies convolutional neural network, and constructs a chemical hepatotoxicity screening model, which saves time, energy and cost, does not involve bioethical issues, and effectively realizes high-throughput screening of chemical hepatotoxicity, providing technical support for filling the gap in chemical hepatotoxicity data.

[0007] The technical means adopted by the present invention are as follows:

[0008] A method for screening hepatotoxicity of chemicals based on high-content cell imaging and convolutional neural networks, comprising:

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

[0010] S2. Randomly select a number of images, use Cellpose2.0 to train a cell nucleus and cell segmentation model suitable for the chemical high-content cell image dataset, use the trained segmentation model to batch process the images, and save the processed Mask images;

[0011] S3, extracting phenotypic parameters based on the saved Mask image, and preprocessing the extracted phenotypic parameters;

[0012] S4. Using the pre-treated phenotypic parameters as input and the presence or absence of hepatotoxicity as the endpoint of the model, a classification model based on a convolutional neural network was constructed;

[0013] S5. The constructed classification model was used for hepatotoxicity screening to obtain the prediction effect.

[0014] Furthermore, step S1 specifically includes:

[0015] S11, select the data numbered idr0088 in the IDR (Image Data Resource) database, refer to the DILIst list, and filter out the high-content cell images of chemicals included in the DILIst list;

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

[0017] Further, in step S11, a total of 34,390 images were screened, which were derived from 400 384-well plates, including 22,400 DMSO-treated images and 11,990 chemical-treated images.

[0018] Further, step S2 specifically includes:

[0019] S21, randomly select 2 images from every 10 plates, a total of 80 images, to train the segmentation model for the chemical high-content cell image dataset;

[0020] S22. Open the visualization interface of Cellpose2.0, select the appropriate basic model, and use the “human-in-the-loop” method to modify the segmented region of interest, so as to train a cell and cell nucleus segmentation model suitable for the dataset;

[0021] S23. Call the trained segmentation model through Python code, perform batch segmentation on the data set, and save the segmented Mask image.

[0022] Further, in step S22, when using Cellpose2.0 to train the cell nucleus segmentation model, 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] Furthermore, step S3 specifically includes:

[0024] S31. Use Python code to extract phenotypic parameters from the mask image, including the area, perimeter, longest axis, shortest axis, roughness, roundness, mean, maximum, minimum, extreme value, variance, median of blue fluorescence intensity, and the number of cells and nuclei, a total of 80 phenotypic parameters;

[0025] S32, preprocessing 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 variances of 0 or close to 0;

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

[0029] S324. Subtract the DMSO background value of the corresponding well plate from the parameter value of the chemical to correct for batch effect. The formula is as follows:

[0030]

[0031] Among them, 0.6745 represents a constant used to convert MAD to a scale close to the standard deviation; X irepresents the i-th data point in the data set; median(X) represents the median of the data set.

[0032] Furthermore, 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 ReLU activation function, each convolutional layer is followed by a 2x2 maximum pooling layer, each pooling layer is followed by a Dropout layer, the Dropout rate is 0.5, the number of samples trained in each batch is 64, and the total number of iterations is 500.

[0033] Furthermore, step S5 specifically includes:

[0034] S51, the data set is divided into a training set and a test set in a ratio of 7:3, and ten-fold cross validation is used for internal validation, and the test set is used for external validation;

[0035] S52, accuracy R using the training set A(train) , Sensitivity R SE(train) Specificity R SP(train) and the area under the receiver operating characteristic curve (ROC) ROC(train) Characterizes the goodness of fit of the model; where:

[0036]

[0037] Among them, R A It indicates the proportion of samples predicted correctly by the model to the total number of samples; R SE It indicates the proportion of samples that are actually positive that are correctly predicted as positive by the model; R SP It represents the proportion of samples that are actually negative that are correctly predicted as negative by the model; TP is true positive, which represents the number of samples correctly predicted as positive by the model; TN is true negative, which represents the number of samples correctly predicted as negative by the model; FP is false positive, which represents the number of samples incorrectly predicted as positive by the model; FN is false negative, which represents the number of samples incorrectly predicted as negative by the model; A ROC It represents the area under the ROC curve, and its value range is between 0 and 1. The closer it is to 1, the better the classification model is at distinguishing between positive and negative classes.

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

[0039] S54. The final prediction effect of the classification model is 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. The present invention provides a method for screening chemical hepatotoxicity based on high-content cell imaging and convolutional neural network. The cell phenotypic parameters obtained by high-content imaging are used as input to construct a convolutional neural network classification model, and internal and external validation is performed. The performance of the model is evaluated by selecting indicators such as accuracy, which can effectively realize 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, provides an effective tool for chemical risk assessment and management, and has important application value.

[0051] Based on the above reasons, the present invention can be widely promoted in the fields of screening of chemical hepatotoxicity risk assessment, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0053] Figure 1 The figure is a flow chart of the method of the present invention.

[0054] Figure 2 A graph showing changes in loss during model training provided by an embodiment of the present invention.

[0055] Figure 3 This is a graph of accuary changes during the model training process provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0056] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work 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 and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0058] like Figure 1 As shown, the present invention provides a method for screening the hepatotoxicity of chemicals based on high-content cell imaging and convolutional neural network, comprising:

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

[0060] S2. Randomly select a number of images, use Cellpose2.0 to train a cell nucleus and cell segmentation model suitable for the chemical high-content cell image dataset, use the trained segmentation model to batch process the images, and save the processed Mask images;

[0061] S3, extracting phenotypic parameters based on the saved Mask image, and preprocessing the extracted phenotypic parameters;

[0062] S4. Using the pre-treated phenotypic parameters as input and the presence or absence of hepatotoxicity as the endpoint of the model, a classification model based on a convolutional neural network was constructed;

[0063] S5. The constructed classification model was used for hepatotoxicity screening to obtain the prediction effect.

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

[0065] S11, select the data numbered idr0088 in the IDR (Image Data Resource) database, refer to the DILIst list (list of drug-induced liver injury classification), and screen out the high-content cell images of chemicals included in the DILIst list;

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

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

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

[0069] S21, randomly select 2 images from every 10 plates, a total of 80 images, to train the segmentation model for the chemical high-content cell image dataset;

[0070] S22. Open the visualization interface of Cellpose2.0, select the appropriate basic model, and use the “human-in-the-loop” method to modify the segmented region of interest, so as to train a cell and cell nucleus segmentation model suitable for the dataset; Figure 2 , 3 As shown in the figure, they are the loss change graph and accuracy change graph during the model training process.

[0071] S23. Call the trained segmentation model through Python code, perform batch segmentation on the data set, and save the segmented Mask image.

[0072] In specific implementation, as a preferred embodiment of the present invention, in step S22, when using Cellpose2.0 to train the cell nucleus segmentation model, the channel is set to [0,0], the cell diameter is 25μm, and when training the cell segmentation model, the channel is set to [3,0], the cell diameter is 37μm.

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

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

[0075] S32, preprocessing 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 variances of 0 or close to 0;

[0078] S323, separate the data of DMSO and chemicals, calculate the median absolute deviation score (MAD_score) of 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), and set the threshold to 3.5. If the MAD_score is greater than 3.5, it is considered as an outlier and removed;

[0079] S324. Subtract the DMSO background value of the corresponding well plate from the parameter value of the chemical to correct for batch effect. The formula is as follows:

[0080]

[0081] Among them, 0.6745 represents a constant used to convert MAD to a scale close to the standard deviation; X i represents the i-th data point in the data set; median(X) represents the median of the data set.

[0082] In specific implementation, as 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 ReLU activation function, each convolutional layer is followed by a 2x2 maximum pooling layer, each pooling layer is followed by a Dropout layer, the Dropout rate is 0.5, the number of samples per batch training (batch_size) is 64, and the total number of iterations (epochs) is 500.

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

[0084] S51, the data set is divided into a training set and a test set in a ratio of 7:3, and ten-fold cross validation is used for internal validation, and the test set is used for external validation;

[0085] S52, accuracy R using the training set A(train) , Sensitivity R SE(train) Specificity R SP(train) and the area under the receiver operating characteristic curve (ROC) ROC(train) Characterizes the goodness of fit of the model; where:

[0086]

[0087]

[0088] Among them, R A It indicates the proportion of samples predicted correctly by the model to the total number of samples; R SE It indicates the proportion of samples that are actually positive that are correctly predicted as positive by the model; R SP It represents the proportion of samples that are actually negative that are correctly predicted as negative by the model; TP is true positive, which represents the number of samples correctly predicted as positive by the model; TN is true negative, which represents the number of samples correctly predicted as negative by the model; FP is false positive, which represents the number of samples incorrectly predicted as positive by the model; FN is false negative, which represents the number of samples incorrectly predicted as negative by the model; A ROC It represents the area under the ROC curve, and its value range is between 0 and 1. The closer it is to 1, the better the classification model is at distinguishing between positive and negative classes.

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

[0090] S54. The final prediction effect of the classification model is 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, it is necessary to screen whether it has hepatotoxicity. First, find the corresponding high-content image in the cell image database or literature, use Cellpose2.0 to segment the cell nucleus and cells, and save the segmented Mask image. Use Python code to extract phenotypic parameters from it, remove missing values, remove phenotypic parameters with variances of 0 or close to 0, remove outliers, and correct batch effects. Then, use the model constructed by the present invention to make predictions. The result is that its activity is 1, which is consistent with the experimental results.

[0101] Example 2

[0102] Given a chemical phenformin, it is necessary to screen whether it has liver toxicity. First, find the corresponding high-content image in the cell image database or literature, use Cellpose2.0 to segment the cell nucleus and cells, and save the segmented Mask image. Use Python code to extract phenotypic parameters from it, remove missing values, remove phenotypic parameters with variances of 0 or close to 0, remove outliers, and correct batch effects. Then use the model constructed by the present invention to make predictions. The result is that its activity is 1, which is consistent with the experimental results.

[0103] Example 3

[0104] Given a chemical amoxapine, it is necessary to screen whether it has hepatotoxicity. First, find the corresponding high-content image in the cell image database or literature, use Cellpose2.0 to segment the cell nucleus and cells, and save the segmented Mask image. Use Python code to extract phenotypic parameters from it, remove missing values, remove phenotypic parameters with variances of 0 or close to 0, remove outliers, and correct batch effects. After these pre-processing operations, use the model constructed by the present invention to predict. The result is that its activity is 0, which is consistent with the experimental results.

[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements 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 network, characterized in that: include: S1. Construct a chemical high-content cell image dataset; S2. Randomly select a number of images, use Cellpose2.0 to train a cell nucleus and cell segmentation model suitable for the chemical high-content cell image dataset, use the trained segmentation model to batch process the images, and save the processed Mask images; S3, extracting phenotypic parameters based on the saved Mask image, and preprocessing the extracted phenotypic parameters; S4. Using the pre-treated phenotypic parameters as input and the presence or absence of hepatotoxicity as the endpoint of the model, a classification model based on a convolutional neural network was constructed; S5. The constructed classification model was used for hepatotoxicity screening to obtain the prediction effect.

2. The method for screening chemical hepatotoxicity based on high-content cell imaging and convolutional neural network according to claim 1, characterized in that: Step S1 specifically includes: S11, select the data numbered idr0088 in the IDR (Image Data Resource) database, refer to the DILIst list, and filter out the high-content cell images of chemicals included in the DILIst list; S12. Label the chemicals with hepatotoxicity as 1 and the chemicals without hepatotoxicity as 0.

3. The method for screening chemical hepatotoxicity based on high-content cell imaging and convolutional neural network according to claim 2, characterized in that: In step S11, a total of 34,390 images were screened, which were derived from 400 384-well plates, including 22,400 DMSO-treated images and 11,990 chemical-treated images.

4. The method for screening chemical hepatotoxicity based on high-content cell imaging and convolutional neural network according to claim 1, characterized in that: Step S2 specifically includes: S21, randomly select 2 images from every 10 plates, a total of 80 images, to train the segmentation model for the chemical high-content cell image dataset; S22. Open the visualization interface of Cellpose2.0, select a suitable basic model, and use the "human-in-the-loop" method to modify the segmented region of interest, thereby training a cell and cell nucleus segmentation model suitable for the dataset; S23. Call the trained segmentation model through Python code, perform batch segmentation on the data set, and save the segmented Mask image.

5. The method for screening chemical hepatotoxicity based on high-content cell imaging and convolutional neural network according to claim 4, characterized in that: In step S22, when using Cellpose2.0 to train the cell nucleus segmentation model, 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.

6. The method for screening chemical hepatotoxicity based on high-content cell imaging and convolutional neural network according to claim 1, characterized in that: Step S3 specifically includes: S31. Use Python code to extract phenotypic parameters from the mask image, including the area, perimeter, longest axis, shortest axis, roughness, roundness, mean, maximum, minimum, extreme value, variance, median of blue fluorescence intensity, and the number of cells and nuclei, a total of 80 phenotypic parameters; S32, preprocessing 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 variances of 0 or close to 0; S323, separate the data of DMSO and chemicals, calculate the median absolute deviation score of each phenotypic parameter in each group, set the threshold to 3.5, and consider the MAD_score greater than 3.5 as an outlier and remove it; S324. Subtract the DMSO background value of the corresponding well plate from the parameter value of the chemical to correct for batch effect. The formula is as follows: Among them, 0.6745 represents a constant used to convert MAD to a scale close to the standard deviation; X i represents the i-th data point in the data set; median(X) represents the median of the data set.

7. The method for screening chemical hepatotoxicity based on high-content cell imaging and convolutional neural network according to claim 1, characterized in that: 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 use ReLU activation function. Each convolutional layer is followed by a 2x2 maximum pooling layer, and each pooling layer is followed by a Dropout layer with a Dropout rate of 0.

5. The number of samples in each batch of training is 64, and the total number of iterations is 500.

8. The method for screening chemical hepatotoxicity based on high-content cell imaging and convolutional neural network according to claim 1, characterized in that: Step S5 specifically includes: S51, the data set is divided into a training set and a test set in a ratio of 7:3, and ten-fold cross validation is used for internal validation, and the test set is used for external validation; S52, accuracy R using the training set A(train) , Sensitivity R SE(train) Specificity R SP(train) and the area under the receiver operating characteristic curve (ROC) ROC(train) Characterizes the goodness of fit of the model; where: Among them, R A It indicates the proportion of samples predicted correctly by the model to the total number of samples; R SE It indicates the proportion of samples that are actually positive that are correctly predicted as positive by the model; R SP It represents the proportion of samples that are actually negative that are correctly predicted as negative by the model; TP is true positive, which represents the number of samples correctly predicted as positive by the model; TN is true negative, which represents the number of samples correctly predicted as negative by the model; FP is false positive, which represents the number of samples incorrectly predicted as positive by the model; FN is false negative, which represents the number of samples incorrectly predicted as negative by the model; A ROC It represents the area under the ROC curve, and its value range is between 0 and 1. The closer it is to 1, the better the classification model is at distinguishing between positive and negative classes. S53, R using the validation set A(test) , R SE(test) , R SP(test) , A ROC(test) Characterize the predictive ability of the model; S54. The final prediction effect of the classification model is as follows: R A(train) ＝0.793 R SE(train) ＝0.794 R SP(train) ＝0.793 A ROC(train) ＝0.881 R A(test) ＝0.671 R SE(test) ＝0.721 R SP(test) ＝0.590 A ROC(test) ＝0.732。

Images