Machine learning-based organelle combination phenotype high-throughput drug screening method

Abstract

The invention discloses an organelle combination phenotype high-throughput drug screening method based on machine learning, and relates to the technical field of drug research and development biology. The method comprises the following steps of: 1, analyzing a cell image by ComOrganNet; step 2, performance evaluation of the ComOrganNet model is carried out; 3, the performance of ComOrganNet in the data set is expanded; step 4, comparing the ComOrganNet with a conventional cell parameter extraction method; 5, verifying the practicability of the ComOrganNet classification model; 6, establishing a natural product library and obtaining an active component screening result; according to the method, cell image analysis and preprocessing are carried out in combination with ComOrganNet, deep learning classification is carried out in combination with OfficientNet V2, higher classification accuracy can be achieved in high-throughput drug screening, particularly, when multi-channel image input is used, the accuracy of a model in a verification set reaches 80%, and the result is remarkably superior to that of single-channel image training.

Description

Technical Field

[0001] The present invention belongs to the field of drug development biotechnology, and in particular relates to a high-throughput drug screening method for organelle combination phenotypes based on machine learning. Background Art

[0002] Traditional drug screening methods usually rely on manual extraction of cell parameters and subjective judgment, which are inefficient and prone to errors. With the development of machine learning technology, automated and high-throughput cell analysis methods have gradually become a research hotspot in the field of drug screening. However, existing machine learning methods still have certain limitations when processing complex cell images and cannot fully tap the potential of cell information. Therefore, it is necessary to propose a new drug screening system to improve screening efficiency and accuracy. Therefore, we propose a high-throughput drug screening method for organelle combination phenotypes based on machine learning. Summary of the invention

[0003] The purpose of the present invention is to provide a high-throughput drug screening method for organelle combination phenotypes based on machine learning. By combining ComOrganNet for cell image analysis and preprocessing, and EfficientNet V2 for deep learning classification, the problem that existing machine learning methods still have certain limitations when processing complex cell images and cannot fully tap the potential of cell information is solved.

[0004] To solve the above technical problems, the present invention is achieved through the following technical solutions: The present invention is a high-throughput drug screening method for organelle combination phenotype based on machine learning, comprising the following steps: Step 1: ComOrganNet cell image analysis process; Step 2: ComOrganNet model performance evaluation; Step 3: ComOrganNet performance in the extended dataset; Step 4: Comparison between ComOrganNet and conventional cell parameter extraction methods; Step 5: Verify the practicality of the ComOrganNet classification model; Step 6: Establishment of the natural product library of Chinese medicine monomer compounds and the results of active ingredient screening.

[0005] Furthermore, the method for high-throughput drug screening of organelle combination phenotypes based on machine learning is characterized in that, in step 1, the ComOrganNet structure described in Section 2.1 is used to identify the center of the cell nucleus and then the image is segmented into 300×300 images ( Figure 2), after cropping, the channel images are merged by the merge function in OpenCV, and then the 300×300×n (n = 3 or 4) images are input into the pre-trained EfficientNet V2 convolutional neural network for classification model training ( Figure 3 ); In order to study the internal logic of classification, an interpretability study was conducted. The images were extracted from the intermediate hidden layers of the EfficientNet V2 convolutional neural network to provide a visual explanation of the model's predictions. We found that most of the shallow units "Crop_Stem conv" and "Crop_Blocks conv" were adjusted to basic collection features such as shape and texture ( Figure 4 ).

[0006] Furthermore, the method for high-throughput drug screening of organelle combination phenotypes based on machine learning is characterized in that, in step 2, we use the EfficientNet V2 convolutional neural network to conduct a binary classification training experiment, and the control group is respectively compared with 100 M Ang-II group, 80 MIS O group, then we trained the neural network in each single channel image, and the loss curve and accuracy curve are shown in Figure 5 and Figure 6 The validation set shown in the figure has low accuracy after 20 epochs, and the loss convergence is not good. The single-channel image performs poorly in the control group-Ang-II group classification and the control group-ISO group classification. Therefore, under the existing model, single-channel images cannot achieve stable prediction; According to the task type, we use 3-channel images for classification, and the convolutional neural network model achieves 80% accuracy in the validation set after 15 epochs ( Figure 7 and Figure 8 ); According to the results of the receiver operating characteristic (ROC) and confusion matrix, the organelle combination - nucleus - lysosome - mitochondria (channels 1, 3, 4) performed best in the binary classification task of the control group and the Ang-II group ( Fig. 9 ); In addition, we use the t-SNE dimensionality reduction method to reduce the dimensionality of the feature vector obtained by the convolution module, and visualize the Fig.10As can be seen from , the feature vector distinguishes the two groups well and reflects the difference between the two groups. In order to quantitatively determine the overlap of the scatter points, we calculated the average silhouette coefficient of the scatter points, where the cluster identifier of each point is its category (predicted category and true category). The scores of the real label scatter plot and the predicted label scatter plot are 2.9% and 3.6% respectively. At the same time, Fig.11 , a similar effect was achieved for the binary classification of the control group-ISO group (AUC=0.99); The scores of the true label scatter plot and the predicted label scatter plot are 4.2% and 4.5% respectively ( Fig.12 ), the results show that ComOrganNet based on combined organelles also has advantages in two binary classification tasks; The purpose of constructing ComOrganNet is to more effectively distinguish the myocardial hypertrophy models constructed by different modeling agents. Therefore, we use ComOrganNet for three-classification training; the training and verification accuracy of similar states of myocardial hypertrophy is as follows Fig.13 As shown; like Fig.14 It shows that the performance of the model trained with single-channel images is poor (the accuracy, recall, and F1 value of each channel are all lower than 0.8). Therefore, single-channel images are not suitable for the classification of the three cardiomyocyte states. like Fig.15 and Fig.16 ,After comparing the image classification models with different channel combinations, we found that the ,model of the mitochondria-lysosome combination achieved higher accuracy (the ,precision, recall rate, and F1 value were 0.91, 0.92, and 0.90, respectively); The result of the confusion matrix ( Fig.17 ), which also confirmed that our model is effective in distinguishing the control group, 100 M Ang-II group, 80 The M ISO group achieved the best results, therefore, in subsequent experiments we will also study mitochondria-lysosomes as combined organelles based on image analysis.

[0007] Furthermore, the machine learning-based high-throughput drug screening method for organelle combination phenotypes is characterized in that, in step three, in addition, this study also used two other extended data sets to verify ComOrganNet based on combined organelles, including a fibrosis cell model image data set (NIH / 3t3 cells and NIH / 3t3 cells modeled with 10 ng / mL TGF-β1, 17760 TIFF format images, such as Fig.18 ), and non-alcoholic fatty liver disease cell model image dataset (HepG2 cells and 200 HepG2 cells modeled by MPA, 6000 TIFF images, such as Fig.19 ); ROC curve and confusion matrix ( Fig. 20 ) shows that ComOrganNet achieved the best results in the above two data sets (AUC=0.99). Similarly, the t-SNE dimension reduction scatter plot of the hidden layer feature vector shows good clustering characteristics. The silhouette coefficients of the real label scatter plot and the predicted label scatter plot in the fibrosis cell model data set are 4.2% and 4.5%, respectively. Fig.21 ), the silhouette coefficients of the true label scatter plot and the predicted label scatter plot in the non-alcoholic fatty liver inflammation cell model dataset are 5.3% and 5.4%, respectively ( Fig. 22 ), these results show that ComOrganNet based on combined organelles has advantages in both datasets of fibrosis cell models and datasets of non-alcoholic steatohepatitis cell models; At the same time, compared with other convolutional neural network model architectures such as VGG19, ResNet50, and DenseNet161, ComOrganNet also shows better performance, including in the three datasets of myocardial hypertrophy, fibrosis, and non-alcoholic fatty liver inflammation cell model. Some classification performance indicators of these network architectures in the three datasets, such as precision, accuracy, recall rate, and F1 value, are shown in Table 23.

[0008] Furthermore, the method for high-throughput drug screening of organelle combination phenotypes based on machine learning is characterized in that, in step 4, the performance of the model is further compared with a previously published conventional method, using two conventional software: Harmony 4.9 and CellProfiler 4.2.6; Section 2.2.4 describes the full details of this part, wherein Harmony 4.9 outputs 147 morphological parameters and CellProfiler 4.2.6 outputs 99 morphological parameters; Then, two morphological parameters were used to perform machine learning of fully connected neural networks and compared with the ComOrganNet method, such as Fig.24 As shown, compared with the conventional methods, ComOrganNet's method performs better than the method with Harmony 4.9 and CellProfiler 4.2.6 output parameters, with higher accuracy, recall, and F1 value (all greater than 0.8); At the same time, the output parameters of the two conventional software were also visualized by t-SNE dimensionality reduction. From the dimensionality reduction scatter plot ( Fig.25) It can be seen that the clustering characteristics of the two software parameters are not obvious, the sample scatter points of different categories overlap with each other, and the prediction results cannot be effectively distinguished. This result intuitively shows that the conventional method cannot effectively complete the task of distinguishing the Ang-II group and the ISO group from the control group; Next, the output parameters of Harmony 4.9 and CellProfiler4.2.6 of the three groups (Ang-II group, ISO group, and control group) were subjected to PCA dimensionality reduction to form a three-dimensional visual scatter plot, which was compared with the output scatter plot of ComOrganNet, as shown in Figure 2. Fig.26 , Harmony 4.9 and CellProfiler 4.2.6 output parameters of each organelle could not distinguish the three groups; Depend on Fig. 27 It can be seen that the output scatter points of ComOrganNet for the three groups of images have clear boundaries, while the limited number of morphological parameters of the two conventional software cannot present clear boundaries between groups. Therefore, ComOrganNet performs better in the assessment of cardiomyocyte hypertrophy without customization or parameter adjustment for any specific dataset; In addition, if Fig.28 As shown in the figure, the feature parameters extracted by traditional commercial software methods cannot achieve effective clustering of the fibrosis cell model dataset and the non-alcoholic fatty liver disease cell model dataset. These results indicate that the number of features extracted by conventional methods is limited and cannot fully extract image information to achieve classification between the control group and the model group. The ComOrganNet model effectively extracts specific features for classification.

[0009] Furthermore, the method for high-throughput drug screening of organelle combination phenotypes based on machine learning is characterized in that, in step 5, in order to verify the practicality of the ComOrganNet classification model, this chapter conducts the following case study: The screening of 105 natural product monomer compounds against cardiomyocyte hypertrophy was used as an example to verify the practicality of ComOrganNet in drug screening.

[0010] Furthermore, the method for high-throughput drug screening of organelle combination phenotypes based on machine learning is characterized in that, in step 6, as shown in Table S2 (Appendix), a natural product library is established from a library of monomeric compounds of traditional Chinese medicine, most of which have a purity of more than 99%; the compound is dissolved in DMSO to prepare a mother solution with a concentration of 10 mM, and stored at -80°C, and a total of 105 compounds are screened at a final concentration of 10 M, and three tests were carried out. A model hole was established on each treatment plate (using 100 After addition of compounds, cells were incubated at 37°C for 24 h. Cells were then fixed and stained with dyes such as HOECHST 33342 as described above. Each compound was tested in triplicate. When analyzed with ComOrganNet, the proportion of normal cells was calculated for each compound, and the 10 compounds with the highest proportions were further investigated.

[0011] Furthermore, the method for high-throughput drug screening of organelle combination phenotypes based on machine learning is characterized in that, in step 6, in order to verify the practicality of ComOrganNet, this study uses ComOrganNet to analyze the cell phenotype images of a homemade natural product library (including 105 Chinese medicine monomer compounds), and based on their effects on 100 The effects of MAng-II cell images were used to identify potential regulators of cardiac hypertrophy. 33×4 images of the nucleus, cytoskeleton, lysosome and mitochondrial channel were obtained by the method described in 3.2.2. These images were collected and then processed by conventional methods and ComOrganNet, respectively. The results showed that 16 compounds were screened out by conventional methods, and 45 compounds were screened out by ComOrganNet, including glycyrrhizin, luteolin-5-O-glucoside, 3'-hydroxypuerarin, 3,4-dimethoxycinnamic acid, 1,3-dicaffeoylquinic acid, astragaloside IV, etc. Among them, these 45 compounds included 16 compounds screened by conventional methods ( Fig.29 , Fig.30 The remaining 45 compounds were further verified and their names and CAS registration numbers are shown in Table S1. It is worth noting that ComOrganNet discovered previously unreported compounds 3,4-dimethoxycinnamic acid and astragaloside II. These results indicate that ComOrganNet has a more accurate screening mode compared with the morphological feature extraction method.

[0012] The present invention has the following beneficial effects: 1. The present invention combines ComOrganNet for cell image analysis and preprocessing, and EfficientNetV2 for deep learning classification, which can achieve higher classification accuracy in high-throughput drug screening. In particular, when using multi-channel image input, the model achieves an accuracy of 80% in the validation set, which is significantly better than the result of single-channel image training. In addition, the mitochondrial-lysosome combined model also shows high precision in the mapping of phenotypic characteristics of mitochondrial targeted compounds, with accuracy, recall rate, and F1 value of 0.91, 0.92, and 0.90, respectively.

[0013] 2. The present invention can provide a visual explanation of the model prediction by analyzing the middle hidden layer images of EfficientNet V2, which enhances the interpretability of the model. Most of the shallow units "Crop_Stem conv" and "Crop_Blocksconv" are adjusted to basic set features such as shape and texture, which helps researchers understand the internal logic and decision-making process of the model. The use of pre-trained EfficientNet V2 models can reduce training time and computing resource consumption while maintaining high performance. ComOrganNet's image processing and analysis capabilities improve data processing efficiency and reduce the need for manual intervention, thereby reducing costs and increasing screening speed.

[0014] 3. The present invention performs well on multiple different data sets, including myocardial hypertrophy, fibrosis and non-alcoholic fatty liver inflammation cell models, showing its good generalization ability. Compared with other convolutional neural network model architectures such as VGG19, ResNet50, and DenseNet161, ComOrganNet also shows better performance, which indicates that the system has broad application prospects and reliability.

[0015] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for describing the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0017] Figure 1 It is a schematic diagram of the process of the high-throughput drug screening method of organelle combination phenotype based on machine learning of the present invention; Figure 2 This is a flowchart of cell image segmentation processing; Figure 3 Classification process for cell images; Figure 4 The cell images are passed through ComOrganNet's internal classification logic; Figure 5 Training accuracy, loss, and ROC curves (single channel) for the control and Ang-II groups; Figure 6 Training accuracy, loss and ROC curves (single channel) for the control group and ISO group; Figure 7 Training accuracy and loss curves for the control group and Ang-II group (multi-channel); Figure 8 Training accuracy and loss curves for the control group and ISO group (multi-channel); Fig. 9 Confusion matrix and ROC curve for the binary classification task of the organelle combination nucleus-lysosome-mitochondria in the control group and the Ang-II group (validation set and test set); Fig.10 t-SNE dimensionality reduction diagram of model feature vector (control group and Ang-II group); Fig.11 Confusion matrix and ROC curve for the binary classification task of the organelle combination nucleus-lysosome-mitochondria in the control group and ISO group (validation set and test set); Fig.12 t-SNE dimensionality reduction diagram of the model feature vector (control group and ISO group); Fig.13 The accuracy curve (top) and loss curve (bottom) of single-channel image three-classification training; Fig.14 The accuracy, recall rate, and F1 value of the single-channel three-classification model; Fig.15 The accuracy curve (top) and loss curve (bottom) of the three-classification training of each channel combination image; Fig.16 The accuracy, recall rate, and F1 value of the three-classification model of each channel; Fig.17 Confusion matrix for the organelle combination nucleus-lysosome-mitochondria in the three-classification task (validation set and test set); Fig.18 Dataset images for fibrosis cell model; Fig.19 Dataset images for non-alcoholic steatohepatitis cell model; Fig. 20 Confusion matrix and ROC curve for the fibrosis and lipid metabolism dataset models; Fig.21 t-SNE dimensionality reduction diagram of the model feature vector (fibrosis dataset); Fig. 22 t-SNE dimensionality reduction diagram of model feature vector (non-alcoholic fatty liver disease cell model dataset); Fig.23 is the classification index in all datasets; Fig.24 The model accuracy, recall rate, and F1 value of the three features; Fig.25 t-SNE dimensionality reduction diagram of Harmony and CellProfiler parameters; Fig.26 PCA dimensionality reduction diagram of single organelle parameters of Harmony (top) and CellProfiler (bottom); Fig. 27 Scatter plots of features extracted for the three processing methods; Fig.28 t-SNE graph of the fibrosis dataset and the non-alcoholic fatty liver disease cell model dataset; Fig.29 A scatter plot covering the screening of 105 natural products (Ang-II); (blue is the median of cell area reduction in conventional methods, n = 3, mean ± SD, and red is the normal cell fraction predicted by ComOrganNet); Fig.30 ComOrganNet and conventional screening results (Ang-II); DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions 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 are within the scope of protection of the present invention.

[0019] See also Figure 1-30 As shown, the present invention is a high-throughput drug screening method for organelle combination phenotype based on machine learning, comprising the following steps: Step 1: ComOrganNet cell image analysis process; Step 2: ComOrganNet model performance evaluation; Step 3: ComOrganNet performance in the extended dataset; Step 4: Comparison between ComOrganNet and conventional cell parameter extraction methods; Step 5: Verify the practicality of the ComOrganNet classification model; Step 6: Establishment of the natural product library of Chinese medicine monomer compounds and the results of active ingredient screening.

[0020] The high-throughput drug screening method for organelle combination phenotype based on machine learning is characterized in that, in step 1, using the ComOrganNet structure in Section 2.1, after identifying the center of the cell nucleus, the image is cropped into a 300×300 segmented image around the center of each cell nucleus ( Figure 2), after cropping, the channel images are merged by the merge function in OpenCV, and then the 300×300×n (n = 3 or 4) images are input into the pre-trained EfficientNet V2 convolutional neural network for classification model training ( Figure 3 ); In order to study the internal logic of classification, an interpretability study was conducted. The images were extracted from the intermediate hidden layers of the EfficientNet V2 convolutional neural network to provide a visual explanation of the model's predictions. We found that most of the shallow units "Crop_Stem conv" and "Crop_Blocks conv" were adjusted to basic collection features such as shape and texture ( Figure 4 ).

[0021] The high-throughput drug screening method for organelle combination phenotype based on machine learning is characterized in that in step 2, we use EfficientNet V2 convolutional neural network to conduct binary classification training experiments, and the control group is respectively compared with 100 M Ang-II group, 80 MIS O group, then we trained the neural network in each single channel image, and the loss curve and accuracy curve are shown in Figure 5 and Figure 6 The validation set shown in the figure has low accuracy after 20 epochs, and the loss convergence is not good. The single-channel image performs poorly in the control group-Ang-II group classification and the control group-ISO group classification. Therefore, under the existing model, single-channel images cannot achieve stable prediction; According to the task type, we use 3-channel images for classification, and the convolutional neural network model achieves 80% accuracy in the validation set after 15 epochs ( Figure 7 and Figure 8 ); According to the results of the receiver operating characteristic (ROC) and confusion matrix, the organelle combination - nucleus - lysosome - mitochondria (channels 1, 3, 4) performed best in the binary classification task of the control group and the Ang-II group ( Fig. 9 ); In addition, we use the t-SNE dimensionality reduction method to reduce the dimensionality of the feature vector obtained by the convolution module, and visualize the Fig.10As can be seen from , the feature vector distinguishes the two groups well and reflects the difference between the two groups. In order to quantitatively determine the overlap of the scatter points, we calculated the average silhouette coefficient of the scatter points, where the cluster identifier of each point is its category (predicted category and true category). The scores of the real label scatter plot and the predicted label scatter plot are 2.9% and 3.6% respectively. At the same time, Fig.11 , a similar effect was achieved for the binary classification of the control group-ISO group (AUC=0.99); The scores of the true label scatter plot and the predicted label scatter plot are 4.2% and 4.5% respectively ( Fig.12 ), the results show that ComOrganNet based on combined organelles also has advantages in two binary classification tasks; The purpose of constructing ComOrganNet is to more effectively distinguish the myocardial hypertrophy models constructed by different modeling agents. Therefore, we use ComOrganNet for three-classification training; the training and verification accuracy of similar states of myocardial hypertrophy is as follows Fig.13 As shown; like Fig.14 It shows that the performance of the model trained with single-channel images is poor (the accuracy, recall, and F1 value of each channel are all lower than 0.8). Therefore, single-channel images are not suitable for the classification of the three cardiomyocyte states. like Fig.15 and Fig.16 ,After comparing the image classification models with different channel combinations, we found that the ,model of the mitochondria-lysosome combination achieved higher accuracy (the ,precision, recall rate, and F1 value were 0.91, 0.92, and 0.90, respectively); The result of the confusion matrix ( Fig.17 ), which also confirmed that our model is effective in distinguishing the control group, 100 M Ang-II group, 80 The M ISO group achieved the best results, therefore, in subsequent experiments we will also study mitochondria-lysosomes as combined organelles based on image analysis.

[0022] The high-throughput drug screening method for combined organelle phenotypes based on machine learning is characterized in that, in step three, in addition, this study also used two other extended data sets to verify ComOrganNet based on combined organelles, including a fibrosis cell model image data set (NIH / 3t3 cells and NIH / 3t3 cells modeled with 10 ng / mL TGF-β1, 17760 TIFF format images, such as Fig.18 ), and non-alcoholic fatty liver disease cell model image dataset (HepG2 cells and 200 HepG2 cells modeled by MPA, 6000 TIFF images, such as Fig.19 ); ROC curve and confusion matrix ( Fig. 20 ) shows that ComOrganNet achieved the best results in the above two data sets (AUC=0.99). Similarly, the t-SNE dimension reduction scatter plot of the hidden layer feature vector shows good clustering characteristics. The silhouette coefficients of the real label scatter plot and the predicted label scatter plot in the fibrosis cell model data set are 4.2% and 4.5%, respectively. Fig.21 ), the silhouette coefficients of the true label scatter plot and the predicted label scatter plot in the non-alcoholic fatty liver inflammation cell model dataset are 5.3% and 5.4%, respectively ( Fig. 22 ), these results show that ComOrganNet based on combined organelles has advantages in both datasets of fibrosis cell models and datasets of non-alcoholic steatohepatitis cell models; At the same time, compared with other convolutional neural network model architectures such as VGG19, ResNet50, and DenseNet161, ComOrganNet also shows better performance, including in the three datasets of myocardial hypertrophy, fibrosis, and non-alcoholic fatty liver inflammation cell model. Some classification performance indicators of these network architectures in the three datasets, such as precision, accuracy, recall rate, and F1 value, are shown in Table 23.

[0023] A high-throughput drug screening method for organelle combination phenotypes based on machine learning, characterized in that, in step 4, the performance of the model is further compared with a previously published conventional method, using two conventional software: Harmony 4.9 and CellProfiler 4.2.6; Section 2.2.4 describes the full details of this part, wherein Harmony 4.9 outputs 147 morphological parameters and CellProfiler 4.2.6 outputs 99 morphological parameters; Then, two morphological parameters were used to perform machine learning of fully connected neural networks and compared with the ComOrganNet method, such as Fig.24 As shown, compared with the conventional methods, ComOrganNet's method performs better than the method with Harmony 4.9 and CellProfiler 4.2.6 output parameters, with higher accuracy, recall, and F1 value (all greater than 0.8); At the same time, the output parameters of the two conventional software were also visualized by t-SNE dimensionality reduction. From the dimensionality reduction scatter plot ( Fig.25) It can be seen that the clustering characteristics of the two software parameters are not obvious, the sample scatter points of different categories overlap with each other, and the prediction results cannot be effectively distinguished. This result intuitively shows that the conventional method cannot effectively complete the task of distinguishing the Ang-II group and the ISO group from the control group; Next, the output parameters of Harmony 4.9 and CellProfiler4.2.6 of the three groups (Ang-II group, ISO group, and control group) were subjected to PCA dimensionality reduction to form a three-dimensional visual scatter plot, which was compared with the output scatter plot of ComOrganNet, as shown in Figure 2. Fig.26 , Harmony 4.9 and CellProfiler 4.2.6 output parameters of each organelle could not distinguish the three groups; Depend on Fig. 27 It can be seen that the output scatter points of ComOrganNet for the three groups of images have clear boundaries, while the limited number of morphological parameters of the two conventional software cannot present clear boundaries between groups. Therefore, ComOrganNet performs better in the assessment of cardiomyocyte hypertrophy without customization or parameter adjustment for any specific dataset; In addition, if Fig.28 As shown in the figure, the feature parameters extracted by traditional commercial software methods cannot achieve effective clustering of the fibrosis cell model dataset and the non-alcoholic fatty liver disease cell model dataset. These results indicate that the number of features extracted by conventional methods is limited and cannot fully extract image information to achieve classification between the control group and the model group. The ComOrganNet model effectively extracts specific features for classification.

[0024] The high-throughput drug screening method for organelle combination phenotype based on machine learning is characterized in that, in step 5, in order to verify the practicality of the ComOrganNet classification model, this chapter conducts the following case study: The screening of 105 natural product monomer compounds against cardiomyocyte hypertrophy was used as an example to verify the practicality of ComOrganNet in drug screening.

[0025] The method for high-throughput drug screening of organelle combinatorial phenotypes based on machine learning is characterized in that, in step 6, as shown in Table S2 (Appendix), a natural product library is established from a library of monomeric compounds of traditional Chinese medicine, most of which have a purity of more than 99%; the compounds are dissolved in DMSO to prepare a mother solution with a concentration of 10 mM and stored at -80°C, and a total of 105 compounds are screened at a final concentration of 10 M, and three tests were carried out. A model hole was established on each treatment plate (using 100 After adding the compounds, the cells were incubated at 37°C for 24 hours. Then, the cells were fixed and stained with dyes such as HOECHST 33342 as before. Each compound was tested in triplicate. When analyzed with ComOrganNet, the proportion of normal cells was calculated for each compound, and the 10 compounds with the highest proportions were further studied.

[0026] The method for high-throughput drug screening of organelle combination phenotypes based on machine learning is characterized in that, in step 6, in order to verify the practicality of ComOrganNet, this study uses ComOrganNet to analyze the cell phenotype images of a self-made natural product library (including 105 Chinese medicine monomer compounds), and based on their effects on 100 The potential regulators of myocardial hypertrophy were identified by the effects of M Ang-II cell images. 33×4 images of the nucleus, cytoskeleton, lysosome and mitochondrial channel were obtained by the 3.2.2 method. These images were collected and then processed by conventional methods and ComOrganNet, respectively. The results showed that 16 compounds were screened out by the conventional method, and 45 compounds were screened out by the ComOrganNet method, including glycyrrhizin, luteolin-5-O-glucoside, 3'-hydroxypuerarin, 3,4-dimethoxycinnamic acid, 1,3-dicaffeoylquinic acid, astragaloside IV, etc. Among them, these 45 compounds included 16 compounds screened by the conventional method ( Fig.29 , Fig.30 The remaining 45 compounds were further verified and their names and CAS registration numbers are shown in Table S1. It is worth noting that ComOrganNet discovered previously unreported compounds 3,4-dimethoxycinnamic acid and astragaloside II. These results indicate that ComOrganNet has a more accurate screening mode compared with the morphological feature extraction method.

[0027] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0028] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to the specific implementation methods described. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A high-throughput drug screening method for organelle combination phenotypes based on machine learning, characterized in that: The steps include: Step 1: ComOrganNet cell image analysis process; Step 2: ComOrganNet model performance evaluation; Step 3: ComOrganNet performance in the extended dataset; Step 4: Comparison between ComOrganNet and conventional cell parameter extraction methods; Step 5: Verify the practicality of the ComOrganNet classification model; Step 6: Establishment of the natural product library of Chinese medicine monomer compounds and the results of active ingredient screening.

2. The method for high-throughput drug screening based on machine learning of organelle combination phenotypes according to claim 1, characterized in that: In step 1, using the ComOrganNet structure described in Section 2.1, after identifying the center of the cell nucleus, the image is cropped with a 300×300 segmentation image around the center of each cell nucleus (Figure 2). After cropping, the channel images are merged by the merge function in OpenCV. Subsequently, the 300×300×n (n=3 or 4) image is input into the pre-trained EfficientNet V2 convolutional neural network for classification model training (Figure 3); In order to study the internal logic of classification, an interpretability study was conducted. The images were extracted from the intermediate hidden layers of the EfficientNet V2 convolutional neural network to provide a visual explanation of the model's predictions. We found that most of the units in the shallow layer, "Crop_Stem conv" and "Crop_Blocks conv", were adjusted to basic collective features such as shape and texture (Figure 4).

3. The method for high-throughput drug screening based on machine learning of organelle combination phenotypes according to claim 1, characterized in that: In step 2, we used the EfficientNet V2 convolutional neural network to conduct a binary classification training experiment. The control group was divided into the 100 μM Ang-II group and the 80 μM ISO group. Subsequently, we trained the neural network in each single channel image. The loss curve and the accuracy curve are shown in Figures 5 and 6. The accuracy of the validation set was not high after 20 epochs, and the loss convergence was not good. The single-channel image performed poorly in the control group-Ang-II group classification and the control group-ISO group classification. Therefore, under the existing model, the single-channel image cannot achieve stable prediction; According to the task type, we used 3-channel images for classification, and the convolutional neural network model achieved 80% accuracy in the validation set after 15 epochs (Figures 7 and 8); According to the results of the receiver operating characteristic (ROC) and confusion matrix, the organelle combination - nucleus - lysosome - mitochondria (channels 1, 3, and 4) performed best in the binary classification task of the control group and the Ang-II group ( Figure 9 ); In addition, we use the t-SNE dimensionality reduction method to reduce the dimension of the feature vector obtained by the convolution module. As can be seen from Figure 10, the feature vector distinguishes the two groups well and reflects the difference between the two groups. In order to quantitatively determine the overlap of the scatter points, we calculated the average silhouette coefficient of the scatter points, where the cluster identifier of each point is its category (predicted category and true category). The scores of the real label scatter plot and the predicted label scatter plot are 2.9% and 3.6%, respectively. At the same time, as shown in Figure 11, the binary classification of the control group-ISO group also achieved a similar effect (AUC=0.99); The scores of the true label scatter plot and the predicted label scatter plot are 4.2% and 4.5% respectively (Figure 12). The results show that ComOrganNet based on combined organelles also has advantages in the two binary classification tasks; The purpose of constructing ComOrganNet is to more effectively distinguish cardiac hypertrophy models constructed by different modeling agents. Therefore, we use ComOrganNet for three-classification training; the training and verification accuracy of similar states of cardiac hypertrophy are shown in Figure 13; As shown in Figure 14, the performance of the model trained with single-channel images is poor (the accuracy, recall, and F1 value of each channel are all lower than 0.8). Therefore, single-channel images are not suitable for classifying the three cardiomyocyte states; As shown in Figures 15 and 16, after comparing the image classification models with different channel combinations, we found that the mitochondria-lysosome combination model achieved higher accuracy (accuracy, recall rate, and F1 value were 0.91, 0.92, and 0.90, respectively); The results of the confusion matrix (Figure 17) also confirmed that our model achieved the best effect in distinguishing the control group, 100μM Ang-II group, and 80μM ISO group. Therefore, in subsequent experiments, we will also study mitochondria-lysosomes as combined organelles based on image analysis.

4. The method for high-throughput drug screening based on machine learning for organelle combination phenotypes according to claim 1, characterized in that: In step three, in addition, this study also used two other extended data sets to verify ComOrganNet based on combined organelles, including a fibrosis cell model image data set (NIH / 3t3 cells and 10 ng / mL TGF-β1 modeled NIH / 3t3 cells, 17760 TIFF format images, as shown in Figure 18), and a non-alcoholic fatty liver disease cell model image data set (HepG2 cells and 200 μM PA modeled HepG2 cells, 6000 TIFF images, as shown in Figure 19); The ROC curve and confusion matrix (Figure 20) show that ComOrganNet achieved the best results in the above two data sets (AUC = 0.99). Similarly, the t-SNE dimension reduction scatter plot of the hidden layer feature vector reflects good clustering characteristics. The silhouette coefficients of the real label scatter plot and the predicted label scatter plot in the fibrosis cell model data set are 4.2% and 4.5%, respectively (Figure 21), and the silhouette coefficients of the real label scatter plot and the predicted label scatter plot in the non-alcoholic fatty liver inflammation cell model data set are 5.3% and 5.4%, respectively (Figure 22). These results show that ComOrganNet based on combined organelles has advantages in both the fibrosis cell model data set and the non-alcoholic fatty liver inflammation cell model data set. At the same time, compared with other convolutional neural network model architectures such as VGG19, ResNet50, and DenseNet161, ComOrganNet also shows better performance, including in the three datasets of myocardial hypertrophy, fibrosis, and non-alcoholic fatty liver inflammation cell model. Some classification performance indicators of these network architectures in the three datasets, such as precision, accuracy, recall rate, and F1 value, are shown in Table 23.

5. The method for high-throughput drug screening based on machine learning of organelle combination phenotypes according to claim 1, characterized in that: In step 4, the performance of the model was further compared with previously published conventional methods using two conventional software: Harmony 4.9 and CellProfiler4.2.6; Section 2.2.4 describes the full details of this part, where Harmony 4.9 outputs 147 morphological parameters and CellProfiler 4.2.6 outputs 99 morphological parameters; Then, the two morphological parameters were used to perform machine learning of the fully connected neural network and compared with the ComOrganNet method. As shown in Figure 25, compared with the conventional method, the ComOrganNet method performed better than the method with the output parameters of Harmony 4.9 and CellProfiler4.2.6, with higher accuracy, recall rate, and F1 value (all greater than 0.8); At the same time, the output parameters of the two conventional software were also visualized by t-SNE dimension reduction. From the dimension reduction scatter plot (Figure 26), it can be seen that the clustering characteristics of the two software parameters are not obvious, the sample scatter points of different categories overlap with each other, and the prediction results cannot be effectively distinguished. This result intuitively shows that the conventional method cannot effectively complete the task of distinguishing the Ang-II group and the ISO group from the control group; Next, the output parameters of Harmony 4.9 and CellProfiler 4.2.6 of the three groups (Ang-II group, ISO group, and control group) were subjected to PCA dimensionality reduction to form a three-dimensional visual scatter plot, which was compared with the output scatter plot of ComOrganNet, as shown in Figure 27. The output parameters of each organelle of Harmony 4.9 and CellProfiler 4.2.6 could not distinguish the three groups; As can be seen from Figure 28, the output scatter points of ComOrganNet for the three groups of images have clear boundaries, while the limited number of morphological parameters of the two conventional software cannot present clear boundaries between groups. Therefore, in terms of cardiomyocyte hypertrophy assessment, ComOrganNet performs better without customization or parameter adjustment for any specific dataset; In addition, as shown in Figure 29, the feature parameters extracted by traditional commercial software methods cannot achieve effective clustering of the fibrosis cell model dataset and the non-alcoholic fatty liver disease cell model dataset. These results show that the number of features extracted by conventional methods is limited and cannot fully extract image information to achieve classification between the control group and the model group. The ComOrganNet model effectively extracts specific features for classification.

6. The method for high-throughput drug screening based on machine learning for organelle combination phenotypes according to claim 1, characterized in that: In step 5, to verify the practicality of the ComOrganNet classification model, this chapter conducted the following case study: The screening of 105 natural product monomer compounds against cardiomyocyte hypertrophy was used as an example to verify the practicality of ComOrganNet in drug screening.

7. The method for high-throughput drug screening based on machine learning for organelle combination phenotypes according to claim 1, characterized in that: In step six, as shown in Table S2 (Appendix), a natural product library was established from a library of monomeric compounds from traditional Chinese medicine, most of which had a purity of more than 99%; the compounds were dissolved in DMSO to prepare a mother solution with a concentration of 10 mM and stored at -80°C. A total of 105 compounds were tested three times at a final concentration of 10 μM. Model wells (treated with 100 μM Ang-II) and blank wells (without compound treatment) were established on each treatment plate to control quality. After adding the compounds, the cells were incubated at 37°C for 24 hours. Then, the cells were fixed and stained with dyes such as HOECHST 33342 as described above. Three replicate wells were tested for each compound; when analyzed using ComOrganNet, the proportion of normal cells was calculated for each compound, and the 10 compounds with the highest proportions were further studied.

8. The method for high-throughput drug screening based on machine learning of organelle combination phenotypes according to claim 1, characterized in that: In step 6, in order to verify the practicality of ComOrganNet, this study used ComOrganNet to analyze the cell phenotype images of a self-made natural product library (including 105 Chinese medicine monomer compounds) and based on their responses to 100 μM Ang-I I cell images were used to identify potential regulators of cardiac hypertrophy. 33×4 images of the nucleus, cytoskeleton, lysosome and mitochondrial channel were obtained by the method described in 3.2.

2. These images were collected and then processed by conventional methods and ComOrganNet, respectively. The results showed that 16 compounds were screened by conventional methods, and 45 compounds were screened by ComOrganNet methods, including glycyrrhizin, luteolin-5-O-glucoside, 3'-hydroxypuerarin, 3,4-dimethoxycinnamic acid, 1,3-dicaffeoylquinic acid, astragaloside I, etc. Among them, these 45 compounds included 16 compounds screened by conventional methods (Figures 33 and 34); the remaining 45 compounds were further verified, and their names and CAS registration numbers are shown in Table S1. It is worth noting that ComOrganNet discovered previously unreported compounds 3,4-dimethoxycinnamic acid and astragaloside II. These results indicate that ComOrganNet has a more accurate screening mode compared with the morphological feature extraction method.

Images