A high-throughput anti-fibrosis drug screening method based on no fluorescence image and application thereof

Abstract

The application discloses a high-throughput anti-fibrosis drug screening method based on a fluorescence-free image and application thereof, relates to the technical field of biological medicine, and aims to obtain a rapid and simple method for large-scale drug screening. In the application, DPC images of NIH / 3T3 cells are collected by a high-content instrument, correlation analysis is performed on morphological parameters obtained by self-analysis software and the expression amount of alpha-SMA protein, 13 morphological parameters highly correlated with the expression amount of alpha-SMA protein are screened out, and the 13 morphological parameters are combined with an artificial intelligence method to be used for large-scale drug screening. The method has the advantages of high throughput, high accuracy, simple operation, and the like, can be used to determine whether a drug has an anti-fibrosis effect, can improve detection efficiency, can be applied to prediction of whether a traditional Chinese medicine monomer has an anti-fibrosis effect, and proves the accuracy and effectiveness of the method as an anti-fibrosis drug screening method, thereby providing important experimental basis and technical support for future drug development and treatment of fibrosis diseases.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of biological medicine, and particularly relates to a high-throughput anti-fibrosis drug screening method based on machine learning and without fluorescence images and application thereof. BACKGROUND

[0002] In recent years, fibrosis problems accompanying the progress of various diseases have caused an increase in fibrous connective tissue in organ tissues, destruction of organ structure, dysfunction, and even failure, which seriously threatens human health and life. At present, there is a lack of effective drugs for treating various types of fibrosis, and new treatment strategies and methods are urgently needed. However, the existing anti-fibrosis screening is only based on fluorescence staining of fibrosis indicators such as proliferation or alpha-SMA protein, which is an expensive and low-throughput screening. In addition, previous researches focus on blocking cell proliferation, rather than maintaining the original state of fibrocytes. Because fibrocytes also play an important role in many tissues and organs. SUMMARY

[0003] The application aims to provide a high-throughput anti-fibrosis drug screening method based on machine learning and without fluorescence images and application thereof. By providing an anti-fibrosis phenotype screening method without special fluorescence labeling, high-throughput, unbiased, and in line with the characteristics of fibrosis diseases, the problem of the existing anti-fibrosis screening being only based on fluorescence staining of fibrosis indicators such as proliferation or alpha-SMA protein, which is an expensive and low-throughput screening, and previous researches focusing on blocking cell proliferation rather than maintaining the original state of fibrocytes is solved.

[0004] To solve the above technical problems, the application is implemented by the following technical scheme:

[0005] The application is a high-throughput anti-fibrosis drug screening method based on machine learning and without fluorescence images and application thereof, which comprises the following steps:

[0006] Step one: collecting DPC imaging and fluorescence imaging under the stimulation of different concentrations of TGF-beta I (0.1 ng / ml, 0.5 ng / ml, 1 ng / ml, 5 ng / ml, 10 ng / ml);

[0007] Step two: collecting DPC imaging parameters and alpha-SMA fluorescence imaging under the same TGF-beta I concentration (10 ng / mL) and different times (3 h, 6 h, 12 h, 24 h, 36 h);

[0008] Step three: SD-208 is a selective TGF-β I inhibitor, this part uses DPC imaging technology to image NIH / 3T3 under different concentrations of SD-208 (50nM, 25nM, 10nM, 5nM, 1nM), and at the same time, immunofluorescence imaging of α-SMA is carried out.

[0009] Step four: construction of high content imaging method;

[0010] Step five: establishment of bright field digital phase contrast imaging (DPC) method;

[0011] Step six: establishment of image analysis method assisted by "machine learning" technology;

[0012] Step seven: detection of 10 natural medicine active compounds against fibrosis.

[0013] 1 Resveratrol 501-36-0

[0014] 2 Matairesinol 580-72-3

[0015] 3 Catechol 120-80-9

[0016] 4 Neo-przewaquinone A 630057-39-5

[0017] 5 Salvianolic acid F 158732-59-3

[0018] 6 Eupalinolide O 2170228-67-6

[0019] 7 Platycodin D 58479-68-8

[0020] 8 Isolinderalactone 957-66-4

[0021] 9 Alantolactone 546-43-0

[0022] 10 Eupalinolide A 877822-40-7

[0023] Configure 1 μM, 5 μM, 10 μM, 20 μM of 10 compounds to act on NIH / 3T3 cells for 24 h.

[0024] Further, the high-throughput anti-fibrosis drug screening method based on fluorescence-free images and the application thereof, characterized in that in step one, 33 morphological parameters are obtained by using Harmony 4.9 software analysis, and the 33 morphological parameters are respectively subjected to correlation analysis with the fluorescence intensity of alpha-SMA, and parameters with a correlation coefficient greater than 0.3 are selected, and through correlation analysis, 24 DPC parameters (r2>0.3) such as Cell Area and Cell Eadial Mean are correlated with the expression of alpha-SMA.

[0025] Further, the high-throughput anti-fibrosis drug screening method based on fluorescence-free images and the application thereof, characterized in that in step two, 33 morphological parameters are obtained by using the Harmony 4.9 software analysis provided by the high-content; 33 morphological parameters are respectively subjected to correlation analysis with the fluorescence intensity of alpha-SMA, and parameters with a correlation coefficient greater than 0.3 are selected, and through correlation analysis, 19 DPC parameters (r2>0.3) such as Cell Area and Cell Eadial Mean are correlated with the expression of alpha-SMA.

[0026] In step three, 33 morphological parameters are obtained by using the Harmony 4.9 software analysis provided by the high-content, and the 33 morphological parameters are respectively subjected to correlation analysis with the fluorescence intensity of alpha-SMA, and parameters with a correlation coefficient greater than 0.3 are selected, and under the influence of SD-208, 16 morphological parameters such as the cell area of DPC imaging cells and cell DPC SER right 0 px are correlated with the fluorescence value (r2>0.3) of alpha-SMA.

[0027] Further, the high-throughput anti-fibrosis drug screening method based on non-fluorescent images and the application thereof are characterized in that, according to the above screening standard, linear correlation analysis is performed on the DPC parameters under the influence of different concentrations of TGF-βI, different time TGF-βI and different concentrations of target inhibitor SD-208, and through analysis of the relationship between different groups of parameters (the correlation coefficient is greater than 0.3), a total of 13 DPC parameters are common disturbance parameters, including Cell Area, Cell Digital Phase Contrast SER Bright 0 px, Cell Digital Phase Contrast SER Hole 0 px, Cell Digital Phase Contrast SER Ridge 0 px, Cell Length, Cell Radial Mean, Cell Roundness, Cell Symmetry 05, Cell Symmetry 15, Cell Threshold Compactness 30%, Cell Threshold Compactness 40%, Cell Threshold Compactness 50%, Cell Threshold Compactness 60%, and these parameters will be used for disturbance mechanism prediction.

[0028] Cell Area: The increase of cell area may indicate that the cell is expanding, which is common in the process of fibroblast activation and fibrosis;

[0029] Cell Eadial Mean: This refers to the radial mean of the cell, reflecting the average distribution of cell size and shape;

[0030] Cell Digital Phase Contrast SER: These parameters may be related to the complexity and irregularity of the cell edge, and may reflect the changes in cell activation and migration ability.

[0031] Further, the high-throughput anti-fibrosis drug screening method based on non-fluorescent images and the application thereof are characterized in that, the experimental design;

[0032] TGF-βI induction experiment;

[0033] The researchers designed a series of experiments to evaluate the effects of different concentrations of TGF-β I on NIH / 3T3 cells, a commonly used mouse fibroblast cell line. By using different concentrations of TGF-β I, the potential effects of this factor under different pathological conditions in vivo can be simulated.

[0034] SD-208 inhibitor experiments;

[0035] As a selective TGF-β I inhibitor, SD-208 can block the signaling of TGF-β I. By observing the changes in cells after SD-208 treatment, a deeper understanding of the role of the TGF-β I signaling pathway in cell morphological and functional changes can be gained.

[0036] Experimental methods;

[0037] Digital phase pair (DPC) imaging;

[0038] DPC imaging technology provides a method for observing cell morphology without the need for fluorescent labeling. This technology is particularly suitable for real-time monitoring of the dynamic process of cell response to various drugs or growth factors.

[0039] α-SMA immunofluorescence imaging;

[0040] α-Smooth muscle actin (α-SMA) is a marker of fibroblast activation, and its expression is usually increased during fibrosis. By immunofluorescence staining, the expression level of α-SMA can be quantitatively analyzed to evaluate the activation state of cells.

[0041] Harmony 4.9 software analysis;

[0042] Harmony 4.9 software is part of high-content analysis (HCA) and is used to extract morphological parameters from DPC and fluorescence imaging. These parameters can be associated with the biological characteristics of cells (such as α-SMA expression) to discover potential morphological markers.

[0043] Further, the high-throughput anti-fibrosis drug screening method based on non-fluorescent images and its application, characterized in that in step four, the high-content imaging method is constructed:

[0044] (1) Based on the processing method of cell morphological analysis software parameters;

[0045] (2) The proposed deep learning and unbiased screening method combines the processing parameters of commercial software with image-based neural network "machine learning" segmentation network, preprocessing module, and classification network to distinguish the activity of compounds;

[0046] Extraction of cell morphological analysis software parameters:

[0047] This section uses Harmony 4.9 software for analysis; the high content imaging system generates 3-channel output images, one of which is bright field imaging, and the other three are fluorescence imaging; counting is performed through the software, and then cell morphology parameters are extracted through the built-in algorithm; the DPC images of NIH / 3T3 cells are analyzed using Harmony 4.9 software, and 33 morphological parameters are obtained; these morphological parameters and a-SMA fluorescence intensity are analyzed for subsequent experiments; linear regression analysis is used in this section to distinguish the correlation between DPC images and a-SMA protein fluorescence values; linear correlation analysis is performed between DPC parameters.

[0048] “Machine learning” data processing:

[0049] In this section, we use a neural network with Inception V4 architecture to classify images, with an initial input of a 16-bit image containing one bright field channel and one DPC channel; we first adjust the size of each channel of the original image and standardize it, then generate a third channel using the average of the existing two channels to form an RGB image, and input a 299x299x3 image into the Inception V4 network; modify the output dimension of the second-to-last fully connected layer to 6 to classify cell images into 6 types; the Inception v4 architecture neural network is pre-trained on the ImageNet object recognition dataset, which contains 1.2 million non-unit objects of 1000 classes, we use the pre-trained convolutional layers, connect them with randomly initialized fully connected layers, and train the resulting network on our own constructed dataset.

[0050] We used TGF-β I as a modeling agent, and added different concentrations of TGF-β I (0.01 μM, 0.10 μM, 1 μM, 5 μM, 10, 20 μM) to the wells containing NIH / 3T3 cells, and the cells treated with different concentrations of TGF-β I were labeled as types 1-5; we defined type 0 and type 1 cells as normal cells, type 2 and type 3 cells as fibrosis progression cells, and type 4 and type 5 cells as fibrosis cells; a dataset containing 24,000 labeled images was established, evenly distributed among the six types; we randomly divided the dataset into training and validation sets with a 4:1 ratio, resulting in 19,200 and 4,800 respectively; the loss function is Softmax_cross_entrophy, and the optimizer is Adam.

[0051] Further, the high-throughput anti-fibrosis drug screening method based on fluorescence-free images and the application thereof are characterized in that, in step four, the DPC images are collected by using the PerkinElmer high-content imaging analysis system Opera Phenix, and the Harmony 4.9 software is used to analyze 13 morphological parameters. The control group, the model group, the 1 μM group, the 5 μM group, the 10 μM group and the 20 μM group are subjected to principal component analysis (PCA) for cluster analysis of the above multi-dimensional parameters. Subsequently, the fluorescence images of α-SMA are collected. It can be seen that the expression of α-SMA is reduced in the 29 kinds of compounds, fully proving the accuracy and effectiveness of the method as an anti-fibrosis screening method.

[0052] In step four, the experimental method is as follows:

[0053] (1) Cell culture: NIH / 3T3 mouse embryonic fibroblasts are purchased from the National Cell Identification Library of the Chinese Academy of Sciences. The NIH / 3T3 cells are inoculated in Dulbecco's Modified Eagle's medium containing 10% newborn calf serum and 1% penicillin and streptomycin. The culture is placed in a 37°C, 5% CO2 incubator for culture.

[0054] (2) Plating and dosing: After the cells are inoculated in the culture dish, the logarithmic growth phase cells are collected and added to the 96-well black transparent bottom culture plate at a concentration of 5000 cells per hole, 100 μL per hole, 37°C, 5% CO2 incubator for 24 h. Different concentrations of the test compounds 1 μM, 5 μM, 10 μM, 20 μM and 10 ng / mL of TGF-β I and the control group are added, and the subsequent operation is carried out after 24 h of culture in a 37°C, 5% CO2 incubator.

[0055] (3) Extraction of cell morphological analysis software parameters: this part uses the Harmony 4.9 software for analysis; the high-content imaging system collects the DPC channel images; the Harmony 4.9 software is used to collect 13 morphological parameters for analysis of the DPC images of the NIH / 3T3 cells; the C group, the model group, the 1 μM group, the 5 μM group, the 10 μM group and the 20 μM group are subjected to principal component analysis (PCA) for cluster analysis of the above multi-dimensional parameters.

[0056] (4) Immunofluorescence staining: NIH / 3T3 cells were fixed with 4% paraformaldehyde for 15 min; then, after washing 3 times in PBS, the cells were permeabilized with 0.1% Triton X-100 for 15 min, and then the samples were blocked with 10% bovine serum albumin for 30 min; removed; placed in the primary antibody liquid, incubated at 4°C overnight; after washing, the samples were incubated with FITC-conjugated secondary antibody liquid at room temperature for 2 h; then the nuclei were stained with DAPI; finally, high-content imaging and analysis system was used for image acquisition; the fluorescence intensity parameters of the a-SMA channel were collected using Harmony 4.9 software;

[0057] Experimental results;

[0058] (1) The principal component analysis (PCA) clustering effect of 13 parameters of 10 compounds can be distinguished from the model group;

[0059] (2) The a-SMA fluorescence intensity of 10 kinds of compounds has significant difference with the model group, which can reduce the expression of a-SMA; it is fully verified that the screening method using 13 parameters is feasible and accurate;

[0060] (3) IC50 toxicity test was performed on the 10 kinds of compounds, and all the compounds had anti-fibrosis effect at non-toxic dose.

[0061] Further, the high-throughput anti-fibrosis drug screening method based on non-fluorescent images and the application thereof are characterized in that, in step five, the bright-field digital phase contrast imaging method is established.

[0062] Microscopy techniques have proven to be essential tools for quantitatively monitoring biological processes with unprecedented spatial resolution; however, current analysis methods have limitations in the number of measurable parameters or image feature extraction; HCS technology, which combines automated fluorescence microscopy and high-throughput automated acquisition technology, has become a powerful tool for studying cell biology and drug development; With the significant advances in computing power, robotics and automation technology, and image detectors, the amount of information obtained by optical microscopy methods has greatly increased, and the application of high-content imaging (HCI) has become more widespread; The effectiveness of HCI in precision medicine has been confirmed in recent studies; Differential phase-contrast (DPC) is a non-interferometric quantitative phase imaging method that uses an asymmetric imaging procedure; Compared with other methods, DPC is experimentally implemented without the need for fluorescent labeling; In DPC, spatially asymmetric imaging methods are used to generate phase gradient images; This technique can observe live cells, analyze bright field images, and does not use any fluorescent dyes, and is non-toxic; However, a machine learning method is needed to extract more information from DPC images; In recent years, several methods have been proposed to predict fluorescent images from transmitted light images, such as bright field images and phase images; Although the above methods show the possibility of using computational methods to convert biomarker images, none of the existing methods are as specific as our work and verify their applicability in HCS applications.

[0063] In this section, the processing parameters of DPC imaging were optimized by comparing the DPC imaging counting method and the fluorescent labeling method of the cell nucleus; Then, the Opera Phenix high-content analysis system equipped with a high-resolution 14-bit charge-coupled device (CCD) camera was used to image each well at 10x magnification; The Harmony 4.9 software was used for analysis, in which 9 pictures were collected per well; The resolution of one image was 1360x1024 pixels, and the corresponding actual physical size was 0.88x0.66 cm; The imaging system produced two-channel output images, one channel for bright field imaging and the other channel for fluorescent imaging; DPC imaging counting and Hoechst 33342 staining counting were compared at 0, 12, 24 and 36 h, respectively; The results showed that there was no significant difference between the two counting methods; Therefore, compared with the traditional Hoechst 33342 cell nucleus staining method, the established method does not need to be labeled.

[0064] One major advantage of phenotypic screening compared to target-based approaches is that the disease-relevant proteins are kept in their natural environment, resulting in more physiologically relevant outcomes; this part, 31 cell morphological parameters were outputted by deep machine deconvolution analysis of DPC imaging through Harmony 4.9 software; in summary, more parameters can be obtained in one screening experiment through deep analysis of image morphology.

[0065] (1) Parameter relationship under different concentrations of TGF-β I stimulation:

[0066] The results showed that under the stimulation of different concentrations of TGF-β I, the DPC imaging parameters and the fluorescence imaging of α-SMA, the cell area, cell roundness, cell width, length and aspect ratio showed the same trend; at the same time, the fluorescence intensity of α-SMA increased with the increase of different concentrations of TGF-β I;

[0067] Through DPC image analysis, 33 morphological parameters of NIH / 3T3 cells were obtained. Through correlation analysis, 24 DPC parameters such as Cell Area, Cell Eadial Mean and others were correlated with α-SMA expression;

[0068] (2) DPC parameter analysis at different times:

[0069] Subsequently, under the same TGF-β I concentration (10 ng / mL), DPC imaging parameters and α-SMA fluorescence imaging were performed at different times; consistent with the above results, DPC imaging parameters showed the same trend in cell area, cell roundness, cell width, cell length and aspect ratio parameters of α-SMA fluorescence imaging at different times; the expression of α-SMA increased with the extension of the administration time;

[0070] Subsequent experiments analyzed these morphological parameters and α-SMA fluorescence intensity; through DPC image analysis, 33 morphological parameters of NIH-3T3 cells were obtained; through correlation analysis, 19 DPC parameters (r2> 0.3) such as Cell Area, Cell Eadial Mean and others were correlated with α-SMA expression;

[0071] In step six, the image analysis method based on "machine learning" technology is established. In this part, we use the neural network with Inception v4 architecture to classify the images; the initial input is a 16-bit image containing a bright field channel and a DPC channel; we first adjust the size of each channel of the original image and standardize it, then generate a third channel with the average value of the existing two channels to form an RGB (red-green-blue) image, and input a 299x299x3 image into the Inception v4 network; modify the output dimension of the penultimate fully connected layer to 6, and divide the cell image into 6 types; the Inception v4 architecture neural network is pre-trained on the ImageNet object recognition dataset, which contains 1.2 million non-unit objects of 1000 categories; we use the pre-trained convolutional layers, connect them with the randomly initialized fully connected layers, and train the resulting network on our own dataset.

[0072] TGF-β I is used as a modeling agent, and different concentrations of TGF-β I (0.01 μM, 0.10 μM, 1 μM, 5 μM, 10, 20 μM) are added to the wells of cultured NIH-3t3 cells. Cells treated with different concentrations of TGF-β I are labeled as types 1-5; we define type 0 and type 1 cells as normal cells, type 2 and type 3 cells as fibrosis progression arrest cells, and type 4 and type 5 cells as fibrosis cells; a dataset containing 24000 labeled images is established, evenly distributed among the six types; we randomly split the dataset with a training-validation ratio of 4:1, obtaining 19200 and 4800 respectively; the loss function is selected as Softmax_cross_entrophy, and the optimizer is selected as Adam.

[0073] After 120 epochs of training, the network loss converges, and the accuracy of the training set and the validation set reaches 95%; then, we test the model on a test set containing 4200 labeled images, which are independent of the training and validation sets performed by different operators; the model performs well in predicting non-training data, with a calculation accuracy of about 92.30%; the above machine learning model is used to distinguish the images of the above 1400 compounds, and 1400x6 wellsx9 fields = 75600 images are processed.

[0074] Further, the high-throughput anti-fibrosis drug screening method based on non-fluorescent images and the application thereof are characterized in that, in step seven, the target inhibitor intervention is:

[0075] SD-208 is a selective TGF-β I inhibitor. This part uses DPC imaging technology to image NIH / 3T3 under different concentrations of SD-208 (50nM, 25nM, 10nM, 5nM, 1nM). At the same time, immunofluorescence imaging of α-SMA is performed; with the increase of the concentration of SD-208, the fluorescence intensity of α-SMA also decreases, indicating that SD-208 affects the expression of α-SMA protein.

[0076] This part also collects the morphological parameters of DPC imaging of NIH / 3T3 cells under the interference of different concentrations of SD-208; under the influence of SD-208, 16 morphological parameters of DPC imaging cells such as cell area and cell DPC SER right 0 px are related to the fluorescence value of α-SMA.

[0077] In summary, we performed linear correlation analysis on DPC parameters under the influence of different concentrations of TGF-β I, different times of the same TGF-β I and target inhibitor SD-208. By analyzing the relationship between parameters in different groups (correlation coefficient greater than 0.3), a total of 13 DPC parameters are common disturbance parameters; these parameters will be used for disturbance mechanism prediction.

[0078] The high-throughput anti-fibrosis drug screening method based on non-fluorescent images according to claim 1 and its application, characterized by DPC image anti-fibrosis degree discrimination based on unbiased automatic deep machine learning:

[0079] In order to further process DPC image data and integrate image recognition methods of image deep learning into our platform; taking TGF-β I as a modeling agent, the images of cells treated with different concentrations of TGF-β I are labeled as type 0-5; we labeled 2000 images for each type. Through rotation to enhance data, a dataset with a final size of 48000 is obtained; the classification network based on Inception V4 is fine-tuned for 48000 images.

[0080] After 120 epochs of training, the network loss converges, and the accuracy of the training set and the validation set reaches 95%; then, we test the model on a test set containing 4200 labeled images, which are independent of the training and validation sets performed by different operators; the model shows high accuracy when predicting non-training data, with a calculation accuracy of about 92.30%; using the above machine learning model to distinguish the images of the above 1400 compounds, 1400x6 wellsx9 fields=75600 images are processed.

[0081] Integration of deep learning and unbiased automatic assisted discovery of anti-fibrosis lead compounds:

[0082] In summary, 33 lead compounds were screened out by morphological parameters, and 65 lead compounds were screened out by unbiased image learning method; 29 lead compounds were the intersection of the two screening modes; the names and CAS numbers of the 29 compounds are listed in Tables 1-7.

[0083] The number of compounds obtained by the image-based screening mode is more than that of the morphological parameter-based screening, which may be due to the fact that the morphological parameter-based screening results excessively rely on phenotypic morphological parameters; compared with the morphological screening mode, the image-based deep learning is an unbiased screening result.

[0084] The above results show that a total of 29 lead compounds can significantly inhibit the abnormal proliferation of fibroblasts. It is important to maintain the original state of fibroblasts, and this result meets the purpose of screening anti-fibrosis lead compounds.

[0085] In addition, the pharmacodynamics of the screening results is reconfirmed; it is verified by immunofluorescence staining that the 29 compounds screened out all have the ability to down-regulate the expression of α-SMA protein; the results show that the 29 lead compounds all inhibit the expression of α-SMA protein, that is, the fluorescence value of α-SMA in the compound group is significantly lower than that in the model group.

[0086] In summary, we quickly identified 29 Chinese medicine compounds from 1400 Chinese medicine compounds; the 29 compounds significantly maintain the morphology of fibroblasts and inhibit the abnormal proliferation of fibroblasts.

[0087] The present application has the following beneficial effects:

[0088] 1、The present application, by correlating the morphological parameters with the expression level of α-SMA, researchers can quickly identify the changes in cells closely related to the fibrosis process, this rapid screening method significantly improves the preliminary screening efficiency of anti-fibrosis drugs, reduces the experimental cost, the traditional fluorescence staining technology is prone to errors in the operation process, and the DPC imaging technology provides a more stable and easy-to-quantify analysis means, reduces the subjectivity and errors in the experimental process, and improves the accuracy and reliability of the data.

[0089] 2、The present application, all 10 kinds of compounds can reduce the expression of alpha-SMA in different degrees, which shows that they have the potential of anti-fibrosis, through high content imaging and analysis system and Harmony 4.9 software, researchers can quickly screen out compounds with potential therapeutic effect, and further understand their specific effects on cell morphology, this method proves its accuracy and effectiveness as an anti-fibrosis drug screening, provides important experimental basis and technical support for future drug development and treatment of fibrosis diseases, in addition, this high-throughput screening method can accelerate the process of new drug discovery and development, and provide new strategies for disease treatment.

[0090] Of course, implementing any product of the present application does not necessarily require all the advantages described above to be achieved simultaneously. BRIEF DESCRIPTION OF DRAWINGS

[0091] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed for the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0092] Fig. 1 is the correlation of DPC parameters and alpha-SMA expression under the stimulation of different concentrations of TGF-beta I;

[0093] Fig. 2 is the correlation of DPC parameters alpha-SMA fluorescence intensity under different time stimulation;

[0094] Figure 3 Linear correlation analysis of DPC parameters under the influence of different concentrations of target inhibitor SD-208

[0095] Figure 4 DPC imaging and alpha-SMA fluorescence imaging under the stimulation of different concentrations of TGF-beta I;

[0096] Figure 5 DPC imaging and alpha-SMA fluorescence imaging under the stimulation of different concentrations of TGF-beta I;

[0097] Figure 6 DPC imaging and alpha-SMA fluorescence imaging under the influence of different concentrations of SD-208;

[0098] Figure 7 The relationship between different groups of parameters;

[0099] Figure 8 Image processing based on unbiased automatic deep machine learning

[0100] Figure 9There are 13 common disturbance parameters;

[0101] Figure 10 Principal component analysis plots for 13 parameters of 10 compounds;

[0102] Figure 11 The graph shows the difference in α-SMA fluorescence intensity between the 10 compounds and the model group;

[0103] Figure 12 Trend chart of IC50 toxicity test results for 10 compounds. Detailed Implementation

[0104] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention.

[0105] Please refer to Figures 1-11. This invention is a high-throughput antifibrotic drug screening method based on fluorescence imaging and its application, which includes the following steps;

[0106] Step 1: Collect DPC imaging and fluorescence imaging under stimulation with different concentrations of TGF-βI (0.1ng / ml, 0.5ng / ml, 1ng / ml, 5ng / ml, 10ng / ml);

[0107] Step 2: At the same TGF-β I concentration (10 ng / mL), DPC imaging parameters and α-SMA fluorescence imaging were acquired at different time points (3 h, 6 h, 12 h, 24 h, 36 h);

[0108] Step 3: SD-208 is a selective TGF-β I inhibitor. In this part, DPC imaging technology was used to image NIH / 3T3 at different concentrations of SD-208 (50nM, 25nM, 10nM, 5nM, 1nM), and immunofluorescence imaging of α-SMA was performed simultaneously.

[0109] Step 4: Construction of high-content imaging methods;

[0110] Step 5: Establishment of the bright-field digital phase contrast imaging (DPC) method;

[0111] Step Six: Establishing an image analysis method assisted by machine learning technology;

[0112] Step 7: Anti-fibrotic test of 10 natural drug active compounds.

[0113] 1 Resveratrol 501-36-0

[0114] 2 Matairesinol (Podocarpus macrophyllus resin 580-72-3)

[0115] 3 Catechol 120-80-9

[0116] 4 Neo-przewaquinone A (New hydroquinone A) 630057-39-5

[0117] 5. Salvianolic acid F (158732-59-3)

[0118] 6 Eupalinolide O 2170228-67-6

[0119] 7 Platycodin D opalactone O 58479-68-8

[0120] 8 Isolinderalactone (957-66-4)

[0121] 9 Alantolactone (546-43-0)

[0122] 10 Eupalinolide A 877822-40-7

[0123] Ten compounds at concentrations of 1 μM, 5 μM, 10 μM, and 20 μM were prepared and applied to NIH / 3T3 cells for 24 hours.

[0124] A high-throughput antifibrotic drug screening method based on fluorescence imaging without the need for fluorescence imaging and its application are characterized in that, in step one, 33 morphological parameters are obtained by analysis using Harmony 4.9 software. Correlation analysis is performed between the 33 morphological parameters and the fluorescence intensity of α-SMA. Parameters with correlation coefficients greater than 0.3 are selected. Through correlation analysis, 24 DPC parameters (r2>0.3), including Cell Area and Cell Eadial Mean, are correlated with α-SMA expression.

[0125] A high-throughput antifibrotic drug screening method based on fluorescence imaging and its application are disclosed. The method is characterized in that, in step two, 33 morphological parameters are obtained using the built-in Harmony 4.9 software. Correlation analysis is performed between these 33 morphological parameters and the fluorescence intensity of α-SMA. Parameters with correlation coefficients greater than 0.3 are selected. Through correlation analysis, 19 DPC parameters (r²>0.3), including Cell Area and Cell Eadial Mean, are correlated with α-SMA expression.

[0126] In step three, the Harmony 4.9 software included in the high-content analysis was used to obtain 33 morphological parameters. Correlation analysis was performed on the 33 morphological parameters with the fluorescence intensity of α-SMA. Parameters with correlation coefficients greater than 0.3 were selected. Under the influence of SD-208, 16 morphological parameters, including cell area of ​​DPC imaging cells and cell DPC SER right 0 px, were correlated with the fluorescence value of α-SMA (r 2>0.3).

[0127] A high-throughput antifibrotic drug screening method based on fluorescence imaging and its application are disclosed. The method is characterized by performing linear correlation analysis on the DPC parameters under the influence of different concentrations of TGF-β1, different time periods of TGF-β1, and different concentrations of the target inhibitor SD-208, according to the above screening criteria. By analyzing the relationships between different groups of parameters (correlation coefficient greater than 0.3), 13 DPC parameters were identified as common perturbation parameters, including: Cell Area, Cell Digital Phase Contrast SER Bright 0px, Cell Digital Phase Contrast SER Hole 0px, Cell Digital Phase Contrast SER Ridge 0px, Cell Length, Cell Radial Mean, Cell Roundness, Cell Symmetry 05, Cell Symmetry 15, Cell Threshold Compactness 30%, Cell Threshold Compactness 40%, Cell Threshold Compactness 50%, and Cell Threshold Compactness 60%. These parameters will be used to predict the perturbation mechanism.

[0128] Cell Area: An increase in cell area may indicate that the cell is expanding, which is common in fibroblast activation and fibrosis.

[0129] Cell Eadial Mean: This refers to the radial mean of cells, reflecting the average distribution of cell size and shape.

[0130] Cell Digital Phase Contrast SER: These parameters may be related to the complexity and irregularity of the cell edge, and may reflect changes in cell activation and migration capabilities.

[0131] A high-throughput antifibrotic drug screening method based on fluorescence imaging without the need for the method and its application, characterized by the experimental design;

[0132] TGF-βI induction experiment;

[0133] Researchers designed a series of experiments to evaluate the effects of different concentrations of TGF-β I on NIH / 3T3 cells (a commonly used mouse fibroblast cell line). By using different concentrations of TGF-β I, the potential role of this factor under different pathological conditions in vivo can be simulated.

[0134] SD-208 inhibitor experiments;

[0135] SD-208, as a selective TGF-βI inhibitor, can block TGF-βI signal transduction. By observing the changes in cells after SD-208 treatment, we can gain a deeper understanding of the role of the TGF-βI signaling pathway in changes in cell morphology and function.

[0136] Experimental methods;

[0137] Digital phase pair (DPC) imaging;

[0138] DPC imaging technology provides a method for observing cell morphology without fluorescent labeling. This technology is particularly suitable for real-time monitoring of the dynamic process of cell response to various drugs or growth factors.

[0139] α-SMA immunofluorescence imaging;

[0140] α-Smooth muscle actin (α-SMA) is a marker of fibroblast activation, and its expression usually increases during fibrosis. Immunofluorescence staining can be used to quantitatively analyze the expression level of α-SMA, thereby assessing the activation status of cells.

[0141] Harmony 4.9 software analysis;

[0142] Harmony 4.9 software is part of High Content Analysis (HCA) and is used to extract morphological parameters from DPC and fluorescence imaging that can be correlated with cellular biological characteristics such as α-SMA expression to discover potential morphological markers.

[0143] A high-throughput antifibrotic drug screening method based on fluorescence imaging without the need for fluorescence imaging and its application, characterized in that, in step four, a high-content imaging method is constructed:

[0144] (1) Processing method based on parameters of cell morphology analysis software;

[0145] (2) The proposed deep learning and unbiased screening method combines the processing parameters of commercial software with image-based neural network “machine learning” segmentation network, preprocessing module and classification network to distinguish the activity of compounds;

[0146] Extraction of parameters from cell morphology analysis software:

[0147] This section uses Harmony 4.9 software for analysis; the high-content imaging system generates 3-channel output images, one channel for bright-field imaging and the other three channels for fluorescence imaging; cell counting is performed using the software, and then cell morphology parameters are extracted using a built-in algorithm; DPC images of NIH / 3T3 cells are analyzed using Harmony 4.9 software to obtain 33 morphological parameters; these morphological parameters and α-SMA fluorescence intensity are analyzed for subsequent experiments; linear regression analysis is used in this section to distinguish the correlation between DPC images and α-SMA protein fluorescence values; linear correlation analysis is performed between DPC parameters.

[0148] Machine Learning Data Processing:

[0149] In this section, we use a neural network based on the Inception V4 architecture to classify images. The initial input is a 16-bit image containing one bright field channel and one DPC channel. We first adjust the size of each channel of the original image and normalize it. Then, we generate a third channel by averaging the two existing channels to form an RGB image. We then input a 299×299×3 image into the Inception V4 network. We modify the output dimension of the penultimate fully connected layer to 6 to classify cell images into 6 types. The Inception v4 architecture neural network is pre-trained on the ImageNet object recognition dataset, which contains 1.2 million non-cell objects in 1000 categories. We use pre-trained convolutional layers, connect them with randomly initialized fully connected layers, and train the network on our own constructed dataset.

[0150] We used TGF-β I as a modeling agent and added different concentrations of TGF-β I (0.01 μM, 0.10 μM, 1 μM, 5 μM, 10, 20 μM) to the wells of NIH / 3T3 cells. The cells treated were labeled as types 1-5. We defined type 0 and type 1 cells as normal cells, type 2 and type 3 cells as cells in the fibrotic process, and type 4 and type 5 cells as fibrotic cells. We built a dataset containing 24,000 labeled images, evenly distributed across the six types. We randomly split the dataset with a 4:1 training-validation ratio, obtaining 19,200 and 4,800 images respectively. The loss function used was Softmax_cross_entrophy, and the optimizer used was Adam.

[0151] A high-throughput antifibrotic drug screening method based on fluorescence imaging without the need for fluorescence imaging and its application are disclosed. The method is characterized by the following step: In step four, DPC images are acquired using the PerkinElmer high-content imaging analysis system Opera Phenix, and 13 morphological parameters are analyzed using the built-in Harmony 4.9 software. Principal component analysis (PCA) is used to cluster the above multidimensional parameters among the control group, model group, 1μM, 5μM, 10μM, and 20μM groups. Subsequently, α-SMA fluorescence images are acquired, showing that all 29 compounds can reduce α-SMA expression, fully demonstrating the accuracy and effectiveness of this method as an antifibrotic screening tool.

[0152] Step four, experimental methods;

[0153] (1) Cell culture: NIH / 3T3 mouse embryonic fibroblasts were purchased from the National Cell Identification Archives of the Chinese Academy of Sciences. NIH / 3T3 cells were seeded in Dulbecco's Modified Eagle's medium containing 10% newborn calf serum and 1% penicillin and streptomycin; and cultured in a 37°C, 5% CO2 incubator.

[0154] (2) Plating and drug delivery: After the cells were seeded in the culture dish, the logarithmic growth phase cells were collected and added to a 96-well black transparent bottom culture plate at a concentration of 5000 cells per well, 100 μL per well. The plates were incubated at 37°C in a 5% CO2 incubator for 24 h. Then, different concentrations of the test compound 1 μM, 5 μM, 10 μM, 20 μM, and 10 ng / mL of TGF-β I and the control group were added. The plates were incubated at 37°C in a 5% CO2 incubator for 24 h before proceeding with subsequent operations.

[0155] (3) Extraction of cell morphology analysis software parameters: This part uses Harmony 4.9 software for analysis; DPC channel images are acquired by a high-content imaging system; 13 morphological parameters selected by Harmony 4.9 software are used to analyze the DPC images of NIH / 3T3 cells; the C group, model group, 1μM, 5μM, 10μM and 20μM groups are clustered by principal component analysis (PCA) to analyze the above multidimensional parameters.

[0156] (4) Immunofluorescence staining: NIH / 3T3 cells were fixed with 4% paraformaldehyde for 15 min; then washed 3 times in PBS, permeabilized with 0.1% Triton X-100 for 15 min, and blocked with 10% bovine serum albumin for 30 min; the cells were then removed and placed in primary antibody solution and incubated overnight at 4°C; after washing, the cells were placed in FITC-conjugated secondary antibody solution and incubated at room temperature for 2 h; the cell nuclei were then stained with DAPI; finally, images were acquired using a high-content imaging and analysis system; the fluorescence intensity parameters of the α-SMA channel were acquired using Harmony 4.9 software;

[0157] Experimental results;

[0158] (1) The principal component analysis (PCA) clustering results of 13 parameters for 10 compounds were all distinguishable from the model group;

[0159] (2) The α-SMA fluorescence intensity of the 10 compounds was significantly different from that of the model group and all of them could reduce the expression of α-SMA; thus fully verifying the feasibility and accuracy of the screening method using 13 parameters.

[0160] (3) IC50 toxicity tests were conducted on these 10 compounds. At non-toxic doses, all compounds showed anti-fibrotic effects.

[0161] A high-throughput antifibrotic drug screening method based on fluorescence imaging without the need for fluorescence imaging and its application, characterized in that, in step five, a bright-field digital phase contrast imaging method is established:

[0162] Microscopic imaging technology has proven to be an essential tool for the quantitative monitoring of biological processes, offering unprecedented spatial resolution; however, current analytical methods are limited in terms of the number of measurable parameters or image feature extraction. High-content imaging (HCS) technology, combining automated fluorescence microscopy and high-throughput automated acquisition technology, has become a powerful tool for cell biology research and drug development. With significant advancements in computing power, robotics, automation, and image detectors, the amount of information obtained through optical microscopy methods has greatly increased, leading to the development of high-content imaging (HCI). The application of HCI is becoming increasingly widespread; the effectiveness of HCI in precision medicine has been confirmed in recent studies; differential phase contrast (DPC) is a non-interferometric quantitative phase imaging method that utilizes asymmetric imaging procedures; compared with most other methods, DPC is experimentally implemented without fluorescent labeling; in DPC, a spatially asymmetric imaging method is used to generate phase gradient images; this technique can observe live cells, analyze bright field images, does not use any fluorescent dyes, and is non-toxic; however, a machine learning method is needed to extract more information from DPC images; in recent years, several methods have been proposed to predict fluorescence images from transmitted light images, such as bright field images and phase images; although the above methods demonstrate the possibility of using computational methods to transform biomarker images, no existing method is as specific as our work and has validated their applicability in HCS applications.

[0163] In this section, the processing parameters of DPC imaging were optimized by comparing the DPC imaging counting method and the nuclear fluorescence labeling method. Then, each well was imaged at 10×magnification using an Opera Phenix high-content analysis system equipped with a high-resolution 14-bit charge-coupled device (CCD) camera. Analysis was performed using Harmony 4.9 software, with 9 images acquired per well. Each image had a resolution of 1360×1024 pixels, corresponding to an actual physical size of 0.88×0.66 cm. The imaging system produced two-channel output images: one channel for bright-field imaging and the other for fluorescence imaging. The DPC imaging count and Hoechst 33342 staining count were compared at 0, 12, 24, and 36 h. The results showed no significant difference between the two counting methods. Therefore, the established method does not require labeling compared to the traditional Hoechst 33342 nuclear staining method.

[0164] Compared to target-based methods, a major advantage of phenotypic screening is that disease-related proteins are preserved in their natural environment, resulting in more physiological translational results; this part, analyzed by Harmony 4.9 software, outputs 31 cell morphology parameters during DPC imaging; in summary, more parameters can be obtained in a single screening experiment through deep analysis of image morphology.

[0165] (1) Parameter relationship under different concentrations of TGF-βI stimulation:

[0166] The results showed that under different concentrations of TGF-β I stimulation, the DPC imaging parameters and α-SMA fluorescence imaging parameters, cell area, cell roundness, cell width, length and aspect ratio showed the same trend; at the same time, the fluorescence intensity of α-SMA increased with the increase of different concentrations of TGF-β I.

[0167] Thirty-three morphological parameters of NIH / 3T3 cells were obtained through DPC image analysis. Correlation analysis revealed that 24 DPC parameters, including Cell Area and Cell Eadial Mean, were correlated with α-SMA expression.

[0168] (2) Analysis of DPC parameters at different times:

[0169] Subsequently, at the same TGF-β I concentration (10 ng / mL), DPC imaging parameters and α-SMA fluorescence imaging were performed at different time points. Consistent with the above results, the DPC imaging parameters showed the same trend in cell area, cell roundness, cell width, cell length, and aspect ratio parameters of α-SMA fluorescence imaging at different time points; α-SMA expression increased with the extension of drug administration time.

[0170] Subsequent experiments analyzed these morphological parameters and α-SMA fluorescence intensity; 33 morphological parameters of NIH-3T3 cells were obtained through DPC image analysis; correlation analysis showed that 19 DPC parameters (r2>0.3), including Cell Area and Cell Eadial Mean, were correlated with α-SMA expression.

[0171] In step six, an image analysis method based on machine learning technology is established. In this part, we use a neural network with the Inception v4 architecture to classify images. The initial input is a 16-bit image containing a bright field channel and a DPC channel. We first adjust the size of each channel of the original image and standardize it. Then, we generate a third channel by averaging the two existing channels to form an RGB (red-green-blue) image. The 299×299×3 image is then input into the Inception v4 network. The output dimension of the penultimate fully connected layer is modified to 6 to classify cell images into 6 types. The Inception v4 architecture neural network is pre-trained on the ImageNet object recognition dataset, which contains 1.2 million non-unit objects in 1000 categories. We use pre-trained convolutional layers, connect them with randomly initialized fully connected layers, and train the network on our own constructed dataset.

[0172] Using TGF-β I as a modeling agent, cells treated with different concentrations of TGF-β I (0.01 μM, 0.10 μM, 1 μM, 5 μM, 10, 20 μM) in the wells of NIH-3t3 cells were labeled as types 1-5. We defined type 0 and type 1 cells as normal cells, type 2 and type 3 cells as cells with arrested fibrosis, and type 4 and type 5 cells as fibrotic cells. A dataset containing 24,000 labeled images was constructed, evenly distributed across the six types. We randomly split the dataset with a 4:1 training-validation ratio, obtaining 19,200 and 4,800 images respectively. The loss function used was Softmax_cross_entrophy, and the optimizer used was Adam.

[0173] After 120 epochs of training, the network loss converged, and the accuracy on both the training and validation sets reached 95%. We then tested the model on a test set containing 4200 labeled images, independent of the training and validation sets operated by different operators. The model performed well in predicting non-training data, with a computational accuracy of approximately 92.30%. Using the aforementioned machine learning model, we distinguished images of the 1400 compounds, processing 1400 × 6 wells × 9 fields of view = 75600 images.

[0174] A high-throughput antifibrotic drug screening method based on fluorescence imaging without the need for the method and its application, characterized in that, in step seven, targeted inhibitor intervention is performed:

[0175] SD-208 is a selective TGF-βI inhibitor. This section uses DPC imaging technology to image NIH / 3T3 cells at different concentrations of SD-208 (50 nM, 25 nM, 10 nM, 5 nM, 1 nM). Simultaneously, immunofluorescence imaging of α-SMA was performed; as the concentration of SD-208 increased, the fluorescence intensity of α-SMA decreased, indicating that SD-208 affects the expression of α-SMA protein.

[0176] This section also collected morphological parameters of NIH / 3T3 cell DPC imaging under different concentrations of SD-208 interference; under the influence of SD-208, 16 morphological parameters, such as cell area and cell DPC SER right 0 px, were correlated with the fluorescence value of α-SMA.

[0177] In summary, we performed linear correlation analysis on DPC parameters under the influence of different concentrations of TGF-β I, the same TGF-β I at different time points, and the target inhibitor SD-208. By analyzing the relationships between different groups of parameters (correlation coefficients greater than 0.3), a total of 13 DPC parameters were identified as common perturbation parameters; these parameters will be used to predict the perturbation mechanisms.

[0178] A high-throughput antifibrotic drug screening method based on fluorescence image-free imaging according to claim 1 and its application, characterized in that the degree of antifibrosis in DPC images is determined based on unbiased automatic deep machine learning:

[0179] To process DPC image data more deeply and integrate deep learning-based image recognition methods into our platform, we used TGF-β I as a modeling agent, labeling cell images treated with different concentrations of TGF-β I as types 0-5. We labeled 2000 images for each type. Data augmentation was performed by rotation, resulting in a final dataset of 48,000 images. A classification network based on Inception V4 was then fine-tuned on these 48,000 images.

[0180] After 120 epochs of training, the network loss converged, and the accuracy on both the training and validation sets reached 95%. We then tested the model on a test set containing 4200 labeled images, independent of the training and validation sets operated by different operators. The model demonstrated high accuracy in predicting non-training data, with a computational precision of approximately 92.30%. Using the aforementioned machine learning model, we distinguished images of the 1400 compounds, processing 1400 × 6 wells × 9 fields of view = 75600 images.

[0181] Integrating deep learning and unbiased automated assisted discovery of anti-fibrosis lead compounds:

[0182] In summary, through two parts of experiments, 33 lead compounds were screened using morphological parameters and 65 lead compounds were screened using unbiased image learning methods; among them, 29 lead compounds were found to be the intersection of the two screening methods; the names and CAS numbers of the 29 compounds are listed in Tables 1-7.

[0183] Image-based screening yielded more compounds than morphological parameter-based screening, possibly because morphological parameter-based screening results are overly dependent on phenotypic morphological parameters; image-based deep learning provides unbiased screening results compared to morphological screening.

[0184] In summary, the results indicate that a total of 29 lead compounds significantly inhibited abnormal fibroblast proliferation. Importantly, these compounds maintained the pristine state of the fibroblasts, a result consistent with our objective of screening anti-fibrotic lead compounds.

[0185] In addition, the efficacy of the screening results was further confirmed; immunofluorescence staining verified that all 29 selected compounds had the ability to downregulate α-SMA protein expression; the results showed that all 29 lead compounds inhibited the expression of α-SMA protein, that is, the fluorescence value of α-SMA in the compound group was significantly lower than that in the model group.

[0186] In summary, we rapidly identified 29 compounds from 1,400 traditional Chinese medicine compounds; these 29 compounds significantly maintained fibroblast morphology and inhibited abnormal fibroblast proliferation.

[0187] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0188] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A high-throughput antifibrotic drug screening method based on fluorescence imaging-free methods, characterized in that: Includes the following steps; Step 1: Collect DPC imaging and fluorescence imaging under different concentrations of TGF-βI stimulation (0.1 ng / ml, 0.5 ng / ml, 1 ng / ml, 5 ng / ml, 10 ng / ml); Step 2: At the same TGF-βI concentration of 10 ng / mL, DPC imaging parameters and α-SMA fluorescence imaging were acquired at different times of 3 h, 6 h, 12 h, 24 h, and 36 h. Step 3: SD-208 is a selective TGF-βI inhibitor. In this part, DPC imaging technology was used to image NIH / 3T3 at different concentrations of SD-208 (50 nM, 25 nM, 10 nM, 5 nM, and 1 nM). At the same time, immunofluorescence imaging of α-SMA was performed. Step 4: Construction of high-content imaging methods; Step 5: Establishment of the bright-field digital phase contrast imaging (DPC) method; Step Six: Establishing an image analysis method assisted by "machine learning" technology; Step 7: Anti-fibrotic detection of 10 natural drug active compounds; 1 Resveratrol 501-36-0 2 Matairesinol Podocarpus resinol 580-72-3 3 Catechol 120-80-9 4 Neo-przewaquinone A (New hydroquinone A) 630057-39-5 5. Salvianolic acid F (158732-59-3) 6 Eupalinolide O 2170228-67-6 7 Platycodin D opalactone O 58479-68-8 8 Isolinderalactone 957-66-4 9 Alantolactone 546-43-0 10 Eupalinolide A 877822-40-7 Ten compounds at concentrations of 1 μM, 5 μM, 10 μM, and 20 μM were prepared and applied to NIH / 3T3 cells for 24 hours. Based on the above screening criteria, linear correlation analysis was performed on the DPC parameters under the influence of different concentrations of TGF-βI, different time periods of TGF-βI, and different concentrations of the target inhibitor SD-208. By analyzing the relationships between different groups of parameters with correlation coefficients greater than 0.3, 13 common perturbation parameters were identified, including: Cell Area, Cell Digital Phase Contrast SER Bright 0px, Cell Digital Phase Contrast SER Hole 0px, Cell Digital Phase Contrast SER Ridge 0px, Cell Length, Cell Radial Mean, Cell Roundness, Cell Symmetry 05, Cell Symmetry 15, Cell Threshold Compactness 30%, Cell Threshold Compactness 40%, Cell Threshold Compactness 50%, and Cell Threshold Compactness 60%. These parameters will be used for perturbation mechanism prediction. Step four involves the construction of high-content imaging methods: (1) Processing method based on parameters of cell morphology analysis software; (2) The proposed deep learning and unbiased screening method combines the processing parameters of commercial software with image-based neural network "machine learning" segmentation network, preprocessing module and classification network to distinguish the activity of compounds. Extraction of parameters from cell morphology analysis software: This section uses Harmony 4.9 software for analysis. The high-content imaging system generates 3-channel output images, one channel for bright-field imaging and the other three channels for fluorescence imaging. Cell counting is performed using the software, and then cell morphology parameters are extracted using a built-in algorithm. Harmony 4.9 software is used to analyze the DPC images of NIH / 3T3 cells, yielding 33 morphological parameters. These morphological parameters and α-SMA fluorescence intensity are analyzed for subsequent experiments. Linear regression analysis is used to distinguish the correlation between DPC images and α-SMA protein fluorescence values. Linear correlation analysis is performed between DPC parameters. Machine learning data processing: A neural network based on the Inception V4 architecture is used to classify images. The initial input is a 16-bit image containing one bright field channel and one DPC channel. First, the size of each channel of the original image is adjusted and normalized. Then, a third channel is generated by averaging the values ​​of the existing two channels to form an RGB image. This 299×299×3 image is then input into the Inception V4 network. The output dimension of the penultimate fully connected layer is modified to 6, classifying the cell images into 6 types. The Inception v4 architecture neural network is pre-trained on the ImageNet object recognition dataset, which contains 1.2 million non-cell objects in 1000 categories. The network is trained on the constructed dataset using pre-trained convolutional layers connected to randomly initialized fully connected layers. Using TGF-βI as a modeling agent, cells treated with different concentrations of 0.01 μM, 0.10 μM, 1 μM, 5 μM, 10 μM, and 20 μM of MTGF-βI in the wells of NIH / 3T3 cells were labeled as types 1-5. Types 0 and 1 were defined as normal cells, types 2 and 3 as cells in the fibrotic process, and types 4 and 5 as fibrotic cells. A dataset containing 24,000 labeled images was constructed, evenly distributed across the six types. The dataset was randomly split with a training-validation ratio of 4:1, resulting in 19,200 and 4,800 images respectively. The loss function used was Softmax_cross_entrophy, and the optimizer used was Adam. In step five, the bright-field digital phase-contrast imaging method is established: In this section, the processing parameters of DPC imaging were optimized by comparing the DPC imaging counting method and the nuclear fluorescence labeling method. Then, each well was imaged at 10×magnification using an Opera Phenix high-content analysis system equipped with a high-resolution 14-bit charge-coupled device camera. Analysis was performed using Harmony 4.9 software, with nine images acquired per well. Each image had a resolution of 1360×1024 pixels, corresponding to an actual physical size of 0.88×0.66 cm. The imaging system produced two-channel output images: one channel for bright-field imaging and the other for fluorescence imaging. DPC imaging counting and Hoechst 33342 staining counting were compared at 0, 12, 24, and 36 h. The results showed no significant difference between the two counting methods; therefore, the established method does not require labeling compared to the traditional Hoechst 33342 nuclear staining method. Compared to target-based methods, a major advantage of phenotypic screening is that disease-related proteins are preserved in the natural environment, resulting in more physiological translational results; 31 cell morphology parameters were obtained during DPC imaging using depth machine deconvolution output by Harmony 4.9 software analysis; more parameters can be obtained in a single screening experiment through depth analysis of image morphology. (1) Relationship of parameters under different concentrations of TGF-βI stimulation: The results showed that under different concentrations of TGF-βI stimulation, the DPC imaging parameters and α-SMA fluorescence imaging parameters, cell area, cell roundness, cell width, length, and aspect ratio showed the same trend; meanwhile, the fluorescence intensity of α-SMA increased with increasing TGF-βI concentration; 33 morphological parameters of NIH / 3T3 cells were obtained through DPC image analysis; and correlation analysis, including cell area and cell eadial... 24 DPC parameters, including Mean, were correlated with α-SMA expression; (2) DPC parameter analysis at different times: Subsequently, at the same TGF-βI concentration of 10 ng / mL, DPC imaging parameters and α-SMA fluorescence imaging were performed at different times; consistent with the above results, the DPC imaging parameters showed the same trend in cell area, cell roundness, cell width, cell length and aspect ratio parameters of α-SMA fluorescence imaging at different times; α-SMA expression increased with the extension of drug administration time; subsequent experiments analyzed these morphological parameters and α-SMA fluorescence intensity; 33 morphological parameters of NIH-3T3 cells were obtained through DPC image analysis; through correlation analysis, 19 DPC parameters, including Cell Area and Cell Eadial Mean, r2>0.3 were correlated with α-SMA expression; in step six, an image analysis method based on "machine learning" technology was established, using Inception The Inception v4 architecture neural network classifies images. The initial input is a 16-bit image containing one bright field channel and one DPC channel. The size of each channel of the original image is adjusted and normalized. Then, a third channel is generated by averaging the existing two channels to form a red-green-blue RGB image. This 299×299×3 image is then input into the Inception v4 network. The output dimension of the penultimate fully connected layer is modified to 6 to classify cell images into 6 types. The Inception v4 architecture neural network is pre-trained on the ImageNet object recognition dataset, which contains 1.2 million non-unit objects in 1000 categories. The network is trained on the constructed dataset using pre-trained convolutional layers connected to randomly initialized fully connected layers. TGF-βI is used as a modeling agent, with 0.01 μM of 0.01 μM of TGF-βI added to the wells of NIH-3t3 cells.Cells treated with different concentrations of TGF-βI (10 μM, 1 μM, 5 μM, 10 μM, and 20 μM) were labeled as types 1-5. Types 0 and 1 were defined as normal cells, types 2 and 3 as cells arrested in the fibrosis process, and types 4 and 5 as fibrotic cells. A dataset containing 24,000 labeled images was constructed, evenly distributed across the six types. The dataset was randomly split with a 4:1 training-validation ratio, resulting in 19,200 and 4,800 images respectively. The loss function used was Softmax_cross_entry. The optimizer used was Adam. After 120 epochs of training, the network loss converged, and the accuracy on both the training and validation sets reached 95%. The model was then tested on a test set containing 4200 labeled images, independent of the training and validation sets operated by different operators. The model performed well in predicting non-training data, with a computational accuracy of approximately 92.30%. Using the aforementioned machine learning model, images of the 1400 compounds were distinguished, processing 1400 × 6 wells × 9 fields of view = 75600 images.

Images