High-throughput drug screening method, device, equipment and medium

Abstract

The invention discloses a high-throughput drug screening method, device, equipment and medium, and relates to the technical field of drug screening, highly abstract semantic information in a collagen gel image is accurately extracted through sampling branches under a deep neural network, and the microstructure of collagen gel can be more accurately identified and analyzed; extracting low-dimensional resolution information in the collagen gel image through an up-sampling branch, combining the highly abstract semantic information and the low-dimensional resolution information into detail information, and segmenting each pixel in the collagen gel image step by step based on the detail information to obtain a collagen gel area; according to the segmentation method combined with the detail information, the collagen gel area can be identified more accurately, influence factors during acquisition of the collagen gel image are counteracted, and the collagen gel area is accurately segmented so as to be better applied to high-throughput drug screening.

Description

Technical Field

[0001] The present invention relates to the technical field of drug screening, and in particular to a high-throughput drug screening method, device, equipment and medium. Background Art

[0002] High Throughput Screening (HTS) refers to a technical system that is based on experimental methods at the molecular and cellular levels, uses microplates as experimental tool carriers, executes the experimental process with an automated operating system, collects experimental result data with sensitive and fast detection instruments, analyzes and processes the experimental data with computers, detects tens of millions of samples at the same time, and supports the operation of the entire system with a corresponding database.

[0003] In high-throughput drug screening, collagen gel technology is widely used to simulate the real extracellular environment, so as to provide a microenvironment that simulates the real human body as much as possible for the complex and diverse interaction process between cells and extracellular matrix. Therefore, it can be used to simulate the extracellular environment for drug screening, which is closer to the effect of drugs in human tissues. As a technical means, it plays an important role in observing the response of cells under the regulation of different drug conditions. The traditional classic collagen gel is used for drug screening and image analysis as follows: a few hundred microliters of collagen are used to fill the entire hole in each well plate. After the collagen gel is formed, a sharp tool is used to split the edge of the collagen gel to separate the gel from the edge of the hole. The collagen gel is photographed at a fixed height angle with a mobile phone or camera, and the edge of the collagen gel is circled by manually processing the image to obtain the morphological change rate of the collagen gel over time.

[0004] However, the traditional collagen gel technology has the following main shortcomings in practice. First, each collagen needs to cover the entire hole, resulting in a huge consumption of collagen and cells, and it is impossible to prepare high-throughput collagen gel experiments. In the process of separating the collagen gel from the edge of the hole, the sharp tools used to split the collagen can easily damage the shape of the collagen gel, and the uniform stability of the initial shape of the collagen cannot be guaranteed, and the success rate is low and unstable; secondly, in the process of taking images, the area of ​​the collagen is too large to use a microscope to obtain the image, so only a mobile phone or camera can be used. In the process of taking pictures with the above equipment, it is impossible to accurately control the height, angle, and light intensity of each picture, and the uniformity and stability of the experimental data acquisition conditions cannot be guaranteed; this leads to large errors in the segmentation of the collagen gel, resulting in low accuracy in collagen gel recognition, and it is difficult to apply it to high-throughput drug screening. Summary of the invention

[0005] The embodiments of the present invention provide a high-throughput drug screening method, device, equipment and medium, which can solve the problem of low recognition accuracy of collagen gel in the prior art.

[0006] The present invention provides a high-throughput drug screening method, comprising the following steps: Acquiring collagen gel image data; Inputting the collagen gel image into a deep neural network, wherein the deep neural network includes an upsampling branch and a downsampling branch; The downsampling branch is used to extract the weight of each pixel position in the feature map corresponding to the collagen gel image to obtain the feature similarity between each pixel and other pixels, and an attention weight matrix is ​​generated. The attention weight matrix is ​​multiplied by the pixel of the feature map to obtain semantic information. The upsampling branch is used to extract the resolution information in the feature map corresponding to the collagen gel image, and the semantic information is fused with the resolution information to obtain the detail information of the collagen gel image. The collagen gel image is segmented based on the detail information of the collagen gel image to obtain the collagen gel area. High-throughput drug screening is achieved based on the collagen gel area.

[0007] Preferably, the step of acquiring collagen gel image data comprises: Using a pipetting workstation to add polydimethylsiloxane (PDMS) to perform hydrophobic treatment on the surface of the well plate; preparing a cell collagen gel suspension, and using the workstation to drop the suspension into the well plate to form a collagen array; placing the array well plate in a constant temperature incubator to form the collagen gel, thereby obtaining a high-throughput collagen gel array; Different types and concentrations of drugs are added to the collagen gel array; at scientifically preset time points based on the characteristics of different cells and drugs, high-content microscopy is used to observe the target cells and drugs, and image data of the time series changes of the collagen gel are obtained to form image data of the collagen gel.

[0008] Preferably, the extraction of semantic information includes: The collagen gel feature information in the image data is extracted using two layers of continuous residual convolution blocks of a deep neural network to obtain an initial feature map A with the same size of H×W×3 as the original image; The initial feature map A is passed through three convolutional layers to obtain three feature maps of size C×H×W, which are matrix B, matrix C, and matrix D, and the sizes of the three feature maps are changed to H×W; Transpose and multiply matrix C and matrix B, and pass the result through the softmax function normalization layer to obtain the spatial attention weight feature map S, the size of which is N×N; Multiply the matrix D and the spatial attention weight feature map S by transposition, adjust the result to a feature map of dimension C×H×W, and multiply this feature map by the adjustment scale coefficient to obtain the attention weight matrix; The attention weight matrix is ​​pixel-multiplied with the initial feature map to obtain a final feature map of size C×H×W, and the semantic information between any two pixels in the image is identified from the final feature map.

[0009] Preferably, after the semantic information is extracted, edge feature information of the initial feature map is extracted through a downsampling branch, including: The Sobel edge feature extraction filters of the X-axis and Y-axis in the downsampling branch are used to extract the edge gradient information in each direction of the initial feature map. The edge gradient information in each direction is fused and extracted through two convolutional layers to obtain preliminary edge features. The preliminary edge features are processed by two consecutive convolutional layers and one batch normalization layer of the edge residual network. Each consecutive layer and each batch normalization layer includes multiple residual blocks. The processed results are segmented and superimposed with these nonlinear information through the Relu function. The results are passed through a convolutional layer for edge feature fusion to obtain edge feature information.

[0010] Preferably, the high-throughput drug screening comprises: The collagen gel area was converted into a binary image; The binary image is analyzed and calculated for area changes in time series, and the calculated area changes are fitted with a drug efficacy effective action curve to obtain a drug effective action concentration curve; The drug effective concentration curve is used to determine the effectiveness of the drug and achieve high-throughput drug screening.

[0011] The embodiment of the present invention also provides a high-throughput drug screening device, comprising: An image acquisition module, used for acquiring collagen gel image data; A feature recognition module, used to input the collagen gel image into a deep neural network, wherein the deep neural network includes an upsampling branch and a downsampling branch; The downsampling branch is used to extract the weight of each pixel position in the feature map corresponding to the collagen gel image to obtain the feature similarity between each pixel and other pixels, and generate an attention weight matrix, and perform pixel product of the attention weight matrix and the feature map to obtain semantic information; The upsampling branch is used to extract the resolution information in the feature map corresponding to the collagen gel image, and fuse the semantic information with the resolution information to obtain the detail information of the collagen gel image; A collagen gel recognition module is used to segment the collagen gel image and obtain the collagen gel area according to the detailed information of the collagen gel image; The drug screening module is used to achieve high-throughput drug screening based on the collagen gel area.

[0012] An embodiment of the present invention further provides an electronic device, including a memory and a processor; The memory is used to store computer programs; The processor is used to implement the steps of a high-throughput drug screening method as described above when executing the computer program stored in the memory.

[0013] An embodiment of the present invention further provides a computer-readable storage medium for storing a computer program, wherein when the computer program is executed by a processor, the steps of a high-throughput drug screening method as described above are implemented.

[0014] The embodiments of the present invention provide a high-throughput drug screening method, device, equipment and medium. Compared with the prior art, the beneficial effects thereof are as follows: The present invention uses a downsampling branch of a deep neural network to accurately extract highly abstract semantic information in collagen gel images, and can more accurately identify and analyze the microstructure of collagen gel; then, the low-dimensional resolution information in the collagen gel image is extracted through an upsampling branch, and the highly abstract semantic information is combined with the low-dimensional resolution information as detail information. Based on the detail information, each pixel in the collagen gel image is segmented step by step to obtain the collagen gel area. This segmentation method combined with detail information can more accurately identify the collagen gel area, offset the influencing factors when obtaining the collagen gel image, and accurately segment the collagen gel area, so as to be better applied to high-throughput drug screening. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic diagram of the overall process of a high-throughput drug screening method provided by an embodiment of the present invention; Figure 2 A schematic diagram of a prepared high-throughput collagen gel array for a high-throughput drug screening method provided by an embodiment of the present invention; Figure 3 A photograph showing the change of the morphology of a collagen gel over time in a high-throughput drug screening method provided by an embodiment of the present invention is a schematic diagram showing the change process of the morphology of the collagen gel; Figure 4 A schematic diagram of the results of accurate identification and segmentation of collagen gel regions by a deep neural network in a high-throughput drug screening method provided in an embodiment of the present invention; Figure 5 A schematic diagram of a drug action concentration curve obtained by statistically analyzing the area of ​​a collagen gel data set accurately segmented by a neural network in a high-throughput drug screening method provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention, so the present invention is not limited by the specific embodiments disclosed below.

[0017] See also Figure 1 The present invention provides a high-throughput drug screening method, comprising the following steps: Step 1: Prepare collagen gel arrays in high throughput and obtain image data of time series changes of collagen gel.

[0018] Construction of a deep neural network for collagen gel region recognition.

[0019] The collagen gel dataset was used to train the deep neural network, and the trained neural network was used to segment the collagen gel area and analyze it to obtain the effective drug concentration curve.

[0020] Step 2: high-throughput preparation of collagen gel array in step 1 and acquisition of data, including: The collagen gel array was prepared with high throughput by controlling a microliter pipetting workstation with pre-written computer program instructions. After the collagen gel was formed, different types of drugs with gradient concentrations were added using the pipetting workstation. At the preset drug-effective time point, a high-content microscope was used to obtain a high-throughput collagen gel image data set for the target cells and drugs.

[0021] Step 3: Build a deep neural network in step 1. The deep neural network includes the following components: Downsampling branch, a downsampling branch for extracting highly abstract semantic information and feature information of collagen gel.

[0022] The upsampling branch, combined with the highly abstract semantic information extracted from the downsampling branch and the upsampling branch that retains the detail information of the original image, is used to accurately classify each pixel in the image and accurately segment the collagen gel area.

[0023] Step 4: The deep neural network in step 1 includes an input feature map extraction module, a spatial position attention module, an edge feature extraction operator module, a three-level downsampling module, a jump connection module, and a three-level upsampling module.

[0024] First, the input raw image data passes through the initial feature map extraction module, and two layers of consecutive residual convolution blocks are used to extract preliminary collagen gel feature information from the image to obtain a feature map with the same size as the original image (H×W×3).

[0025] In each level of downsampling module and upsampling module, a spatial position attention module and an edge feature extraction module are included.

[0026] The spatial position attention module is used to capture the spatial dependency information between any two pixels in the feature map, and the weight of each pixel position is used to represent the feature similarity between the two positions. During the training process, the module gradually tends to focus on the key area in the image, which is the collagen gel area in the present invention.

[0027] First, the initial feature map A (C×H×W) is first passed through three convolutional layers to obtain three feature maps {B, C, D}, all of which have a size of (C×H×W). They are then transformed into (C×N)N=H×W.

[0028] Secondly, the transpose of matrix C and matrix B is multiplied and the spatial attention weight feature map S (N×N) is obtained through the softmax normalization layer.

[0029] Subsequently, the transpose of the matrices D and S are multiplied together, and the result is resized to obtain an output of (C×H×W), which is then multiplied by the rescaling factor.

[0030] Finally, the obtained attention weight coefficient is used to perform pixel-wise multiplication with the original feature map to obtain the output of this module.

[0031] The edge feature extraction module first sets the Sobel edge feature extraction filters in the X and Y directions to extract the gradient information about the edge in each direction, and then further extracts the edge features after fusing the information through two convolutional layers.

[0032] The initially extracted edge features are then passed through an edge residual network, which consists of two consecutive convolutional layers and a batch normalization layer. Nonlinear information is introduced through the Relu function, and then feature fusion is performed through another convolutional layer.

[0033] In the upsampling branch, the skip connection module is used to combine the high-dimensional semantic information of the collagen gel with the low-dimensional resolution information. The detail information extracted by the above partial neural network is used to identify and segment the collagen gel area step by step in each level of the deconvolution module. Finally, after three upsamplings, the accurate recognition result of the collagen gel area is obtained.

[0034] Step 5: The step of training the deep neural network using the precisely prepared collagen gel dataset; optimizing the parameters of each layer in the deep neural learning network.

[0035] Step 6: Use the data set in step 5 to optimize the parameters to be optimized in each layer of the deep learning network, including: The collagen gel data training set was used to train the deep learning neural network, the collagen gel data validation set was used to verify the performance and convergence of the deep learning network training, and the test data set was used for the final test of the deep learning network.

[0036] Specify the loss function, use AdamW as the optimizer of the model loss, set the learning rate of the cosine annealing strategy, use mFscore as the evaluation criterion of the model, stop the training when the loss value of the entire deep neural network is less than the specified threshold or reaches the set maximum number of iterations, save all the parameters of the deep neural network, merge the training set and the validation set using the best hyperparameter combination, retrain the model, and obtain the best model weights for subsequent inference segmentation of collagen gel.

[0037] Step 7: In step 2, the image data of collagen gel under the action of different types and gradient concentrations of drugs are obtained, and the trained deep neural network is used for inference segmentation to output a binary image obtained by accurate collagen gel area identification.

[0038] Step 8: In step 2, the area change of the time series of the accurately identified collagen gel binary image is calculated, and the effective action curve of the drug is fitted. The effectiveness of the drug can be determined by comparing the efficacy action curves of different drugs. Quantitative analysis of the drug action curve can obtain the effective action concentration of the drug on the collagen gel scale, so that high-throughput screening of drugs can be performed.

[0039] like Figure 1 FIG. 1 is a schematic diagram of the overall process of an embodiment of the present invention, which specifically includes: P1: Use a pipetting workstation to add PDMS or F127 to make the surface of the well plate superhydrophobic.

[0040] P2: Prepare cell collagen gel suspension, use a pipetting workstation to prepare collagen array, place it in a constant temperature incubator to shape the collagen gel, and then add culture medium containing different types and concentrations of drugs to make the collagen gel freely suspended in the culture medium; Figure 2 Shown is the prepared high-throughput collagen gel array.

[0041] P3: According to the pre-set observation time nodes for specific cells and drugs, the data sets of the collagen gel array at different time nodes are obtained; Figure 3The photographs shown are schematic diagrams of the morphological changes of collagen gel over time.

[0042] P4: Use deep learning neural networks to accurately identify and segment collagen gels, such as Figure 4 As shown, the morphological change curve of the collagen gel array under the action of different types and concentrations of drugs is obtained based on the calculated area fitting, as shown in Figure 5 shown in Figure 5 In the figure, the abscissa Log Concentration (uM) represents the logarithmic concentration (uM); the ordinate Collagenshrink rate represents the collagen shrinkage rate; and DOX Collagen Assay represents the DOX collagen assay.

[0043] From a technical point of view, the high-throughput collagen gel array is fast and stable, and truly realizes the high throughput of collagen gel for large-scale drug screening. Unlike the traditional classic two-step method, each collagen droplet consumes very little collagen, which greatly reduces the threshold of drug screening in terms of cost, and takes into account both robustness and flexibility. From the perspective of the entire process, the collagen array is printed using a high-throughput microliter pipetting workstation, and images are taken using a high-content microscope. Deep learning technology is used to quickly analyze images of collagen gel morphological changes. The trained neural network can be highly sensitive to the characteristics of collagen gel, and at the same time, it can extract highly specific visual features of other non-collagen gel areas under the microscope, truly realizing the high-throughput and high-precision characteristics of collagen gel for large-scale drug screening, so that the overall solution of the present invention truly realizes the full-process high-throughput characteristics of collagen gel-based drug screening.

[0044] The present invention prepares a well plate with a super-hydrophobic culture interface in advance, then quickly and stably simulates a collagen gel array of a real extracellular three-dimensional environment in the well plate, captures and collects time series pictures of collagen gel containing cells under the action of different types and gradient concentrations of drugs, and annotates and generates a collagen gel data set for training a deep learning model; constructs a deep learning neural network that can identify the collagen gel area with high precision, and uses the established data set to train the neural network model, and accurately identifies the collagen gel area for the high-throughput data obtained in subsequent experiments; analyzes the collagen gel morphological changes in the time series to obtain a drug action curve, thereby determining the effectiveness of the drug and accurately calculating the effective concentration range of the drug, thereby achieving the purpose of high-throughput drug screening; the present invention uses high-throughput equipment to stably and efficiently prepare a collagen gel suspension array containing cells, obtains collagen gel image data at preset time nodes, uses a deep neural network to identify the collagen gel area, and uses the collagen gel morphological change rate as a phenotype reading, thereby realizing the function of high-throughput drug screening on the basis of ensuring robustness and flexibility.

[0045] An embodiment of the present invention provides a high-throughput drug screening device, comprising: The image acquisition module is used to acquire collagen gel image data.

[0046] The feature recognition module is used to input the collagen gel image into a deep neural network, wherein the deep neural network includes an upsampling branch and a downsampling branch.

[0047] The downsampling branch is used to extract the weight of each pixel position in the feature map corresponding to the collagen gel image to obtain the feature similarity between each pixel and other pixels, and generate an attention weight matrix, and perform pixel multiplication of the attention weight matrix and the feature map to obtain semantic information.

[0048] The upsampling branch is used to extract the resolution information in the feature map corresponding to the collagen gel image, and fuse the semantic information with the resolution information to obtain the detail information of the collagen gel image.

[0049] The collagen gel recognition module is used to segment the collagen gel image and obtain the collagen gel area according to the detailed information of the collagen gel image.

[0050] The drug screening module is used to achieve high-throughput drug screening based on the collagen gel area.

[0051] An embodiment of the present invention provides an electronic device, including a memory and a processor.

[0052] The memory is used to store computer programs.

[0053] The processor is used to implement the steps of the above-mentioned high-throughput drug screening method when executing the computer program stored in the memory.

[0054] An embodiment of the present invention further provides a computer-readable storage medium for storing a computer program, which implements the steps of the above high-throughput drug screening method when executed by a processor.

[0055] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A high-throughput drug screening method, characterized in that: The following steps are involved: Acquiring collagen gel image data; Inputting the collagen gel image into a deep neural network, wherein the deep neural network includes an upsampling branch and a downsampling branch; The downsampling branch is used to extract the weight of each pixel position in the feature map corresponding to the collagen gel image to obtain the feature similarity between each pixel and other pixels, and an attention weight matrix is ​​generated. The attention weight matrix is ​​multiplied by the pixel of the feature map to obtain semantic information. The upsampling branch is used to extract the resolution information in the feature map corresponding to the collagen gel image, and the semantic information is fused with the resolution information to obtain the detail information of the collagen gel image. The collagen gel image is segmented based on the detail information of the collagen gel image to obtain the collagen gel area. High-throughput drug screening is achieved based on the collagen gel area.

2. A high-throughput drug screening method according to claim 1, characterized in that: The step of acquiring collagen gel image data comprises: Using a pipetting workstation to add polydimethylsiloxane (PDMS) to perform hydrophobic treatment on the surface of the well plate; preparing a cell collagen gel suspension, and using the workstation to drop the suspension into the well plate to form a collagen array; placing the array well plate in a constant temperature incubator to form the collagen gel, thereby obtaining a high-throughput collagen gel array; Different types and concentrations of drugs are added to the collagen gel array; at scientifically preset time points based on the characteristics of different cells and drugs, high-content microscopy is used to observe the target cells and drugs, and image data of the time series changes of the collagen gel are obtained to form image data of the collagen gel.

3. A high-throughput drug screening method according to claim 1, characterized in that: The extraction of the semantic information includes: The collagen gel feature information in the image data is extracted using two layers of continuous residual convolution blocks of a deep neural network to obtain an initial feature map A with the same size of H×W×3 as the original image; The initial feature map A is passed through three convolutional layers to obtain three feature maps of size C×H×W, which are matrix B, matrix C, and matrix D, and the sizes of the three feature maps are changed to H×W; Transpose and multiply matrix C and matrix B, and pass the result through the softmax function normalization layer to obtain the spatial attention weight feature map S, the size of which is N×N; Multiply the matrix D and the spatial attention weight feature map S by transposition, adjust the result to a feature map of dimension C×H×W, and multiply this feature map by the adjustment scale coefficient to obtain the attention weight matrix; The attention weight matrix is ​​pixel-multiplied with the initial feature map to obtain a final feature map of size C×H×W, and the semantic information between any two pixels in the image is identified from the final feature map.

4. A high-throughput drug screening method according to claim 3, characterized in that: After the semantic information is extracted, edge feature information of the initial feature map is extracted through a downsampling branch, including: The Sobel edge feature extraction filters of the X-axis and Y-axis in the downsampling branch are used to extract the edge gradient information in each direction of the initial feature map. The edge gradient information in each direction is fused and extracted through two convolutional layers to obtain preliminary edge features. The preliminary edge features are processed by two consecutive convolutional layers and one batch normalization layer of the edge residual network. Each consecutive layer and each batch normalization layer includes multiple residual blocks. The processed results are segmented and superimposed with these nonlinear information through the Relu function. The results are passed through a convolutional layer for edge feature fusion to obtain edge feature information.

5. A high-throughput drug screening method according to claim 1, characterized in that: The high-throughput drug screening method comprises: The collagen gel area was converted into a binary image; The binary image is analyzed and calculated for area changes in time series, and the calculated area changes are fitted with a drug efficacy effective action curve to obtain a drug effective action concentration curve; The drug effective concentration curve is used to determine the effectiveness of the drug and achieve high-throughput drug screening.

6. A high-throughput drug screening device, characterized in that: include: An image acquisition module, used for acquiring collagen gel image data; A feature recognition module, used to input the collagen gel image into a deep neural network, wherein the deep neural network includes an upsampling branch and a downsampling branch; The downsampling branch is used to extract the weight of each pixel position in the feature map corresponding to the collagen gel image to obtain the feature similarity between each pixel and other pixels, and generate an attention weight matrix, and perform pixel product of the attention weight matrix and the feature map to obtain semantic information; The upsampling branch is used to extract the resolution information in the feature map corresponding to the collagen gel image, and fuse the semantic information with the resolution information to obtain the detail information of the collagen gel image; A collagen gel recognition module is used to segment the collagen gel image and obtain the collagen gel area according to the detailed information of the collagen gel image; The drug screening module is used to achieve high-throughput drug screening based on the collagen gel area.

7. An electronic device, characterized in that: include: Memory and processor; The memory is used to store computer programs; The processor is used to implement the steps of a high-throughput drug screening method as described in any one of claims 1 to 5 when executing the computer program stored in the memory.

8. A computer-readable storage medium, characterized in that: Used to store a computer program, which, when executed by a processor, implements the steps of a high-throughput drug screening method as described in any one of claims 1 to 5.

Images