Chromosome conformation contact map enhancement method and apparatus based on generative adversarial network

Abstract

The present application relates to the technical field of chromosome sequencing enhancement, in particular to a chromosome conformation contact map enhancement method and device based on a generative adversarial network. The present application fully considers the correlation between base distance and chromosome conformation contact map, inputs the base distance as condition information into the generative adversarial network, so that the generative adversarial network can distinguish between chromosome conformation contact maps with different base distances during training, and then the model after training can enhance the resolution of each chromosome conformation contact map according to the base distance, thereby improving the overall resolution of the chromosome conformation contact map.

Description

Technical Field

[0001] This invention relates to the field of chromosome sequencing enhancement technology, specifically to a method and apparatus for enhancing chromosome conformation contact maps based on generative adversarial networks. Background Technology

[0002] Human chromosomes possess a complex, multi-level, three-dimensional spatial structure. Precisely analyzing this structure helps reveal the mechanisms of gene regulation, thereby enabling the prevention and treatment of genetic diseases. High-throughput chromosome conformation capture (Hi-C) technology is currently the primary method for obtaining chromosome three-dimensional structures. The common practice is to construct chromosome conformation contact maps based on Hi-C sequencing data, and then further analyze and statistically analyze these maps to determine the presence of chromosomal abnormalities. High-resolution Hi-C sequencing data requires significant human, material, and time investment, but the chromosome conformation contact maps constructed from high-resolution Hi-C sequencing data can accurately reveal the three-dimensional structure of chromosomes (i.e., high-resolution chromosome conformation contact maps can accurately reveal the three-dimensional structure of chromosomes). Low-resolution Hi-C sequencing data offers advantages in speed and cost, but chromosome conformation contact maps constructed from low-resolution Hi-C sequencing data cannot accurately reveal the three-dimensional structure of chromosomes and may sometimes even lead to erroneous conclusions. However, most available Hi-C data currently have a resolution of 25 to 100 Mb. Therefore, there is a need to develop methods to improve the resolution of Hi-C contact maps (chromosome conformation contact maps). Existing Hi-C contact map enhancement methods include those based on Convolutional Neural Networks (CNNs). CNN-based contact map enhancement methods typically use mean squared error as the loss function, resulting in over-smoothing, blurred edges, and strong distance dependence in the enhanced contact maps. For example, HiCPlus, a HiC contact map enhancement method based on CNNs proposed in 2018, produces highly blurred edges of topological domains and chromatin loops in the enhanced contact map, which is detrimental to downstream bioinformatics analysis. Furthermore, because CNN-based chromosome conformation contact map enhancement methods do not consider base distances, the quality of the enhanced chromosome conformation contact maps is poor.

[0003] In summary, existing enhancement methods result in poor quality chromosome conformation contact maps.

[0004] Therefore, existing technologies still need to be improved and enhanced. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method and apparatus for enhancing chromosome conformation contact maps based on generative adversarial networks, which solves the problem of poor quality chromosome conformation contact maps caused by existing enhancement methods.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a method for enhancing chromosome conformation contact maps based on generative adversarial networks, comprising:

[0008] Based on the location of each region in the chromosome in the original chromosome conformation contact map, the original chromosome conformation contact map is processed to obtain the sample chromosome conformation contact map;

[0009] Based on the contact map of the sample chromosome conformation and the base distance of the contact map of the sample chromosome conformation, a generative adversarial network is trained to obtain the trained generative adversarial network.

[0010] The contact map of the chromosome conformation to be enhanced is input into the trained generative adversarial network to obtain the enhanced chromosome conformation contact map output by the trained generative adversarial network.

[0011] In one implementation, the step of processing the original chromosome conformation contact map based on the positions of each region within the chromosome to obtain a sample chromosome conformation contact map includes:

[0012] The pixel values ​​of the original chromosome conformation contact map are normalized to obtain the normalized original chromosome conformation contact map.

[0013] The normalized original chromosome conformation contact map is divided into blocks according to a set size to obtain sample sub-maps that characterize each region of the original chromosome conformation contact map.

[0014] Remove the sample sub-images located below the diagonal in each of the sample sub-images, as well as the sample sub-images located at the centromere of the chromosome, to obtain the pre-selected sample sub-images;

[0015] Based on the base distances of the pre-selected sample sub-images, the pre-selected sample sub-images are screened to obtain sample chromosome conformation contact maps.

[0016] In one implementation, the step of filtering the pre-selected sample sub-images based on the base distances of the pre-selected sample sub-images to obtain a sample chromosome conformation contact map includes:

[0017] After removing the sample sub-images whose base distance is greater than or equal to a set distance from each of the pre-selected sample sub-images, a sample chromosome conformation contact map is obtained.

[0018] In one implementation, training a generative adversarial network based on the contact map of the sample chromosome conformation and the base distance between the contact maps of the sample chromosome conformation to obtain the trained generative adversarial network includes:

[0019] Based on the contact map of the sample chromosome conformation and the base distance of the contact map of the sample chromosome conformation, the generator and discriminator to be trained are trained alternately to obtain the trained generative adversarial network, and the generator and the discriminator constitute the generative adversarial network.

[0020] In one implementation, the generator and discriminator are alternately trained based on the base distance between the sample chromosome conformation contact map and the sample chromosome conformation contact map to obtain the trained generative adversarial network, wherein the generator and the discriminator constitute the generative adversarial network, including:

[0021] The contact map of the sample chromosome conformation and the base distance of the contact map of the sample chromosome conformation are input into the generator to be trained to obtain the enhanced resolution sample map output by the generator to be trained.

[0022] The enhanced resolution sample image and the true set resolution image output by the generator to be trained are input into the discriminator to be trained to obtain the generation loss function output by the discriminator to be trained. The generation loss function is used to characterize the degree of difference between the enhanced resolution sample image and the true set resolution image.

[0023] Based on the generation loss function, the parameters of the generator to be trained are adjusted to complete one round of training of the generator;

[0024] The sample chromosome conformation contact map and the base distance of the sample chromosome conformation contact map are input into the generator after one round of training to obtain the enhanced resolution sample map output by the generator after one round of training.

[0025] The enhanced resolution sample image output by the generator after one round of training and the real set resolution image are input into the discriminator to be trained to obtain the output value of the discriminator to be trained.

[0026] Calculate the true loss value of the enhanced resolution sample image output by the generator after one round of training, relative to the true set resolution image;

[0027] Calculate the loss value of the output value of the discriminator to be trained relative to the output value of the desired discriminator, denoted as the discriminant loss function;

[0028] Based on the discriminant loss function, the parameters of the discriminator to be trained are adjusted to obtain one round of training for the discriminator;

[0029] Based on the generator and discriminator trained in one round, the generator and discriminator are trained alternately until the number of iterations reaches the set number, thus completing the training of the generative adversarial network.

[0030] In one implementation, the step of inputting the contact map of the chromosome conformation to be enhanced into the trained generative adversarial network to obtain the enhanced chromosome conformation contact map output by the trained generative adversarial network includes:

[0031] The contact map of the chromosome conformation to be enhanced is divided into blocks according to a set size to obtain each sub-map to be enhanced;

[0032] Select the subgraph to be enhanced located at the diagonal position of the chromosome from each of the subgraphs to be enhanced, and denote it as the first subgraph;

[0033] Select the subgraph to be enhanced located below the diagonal of the chromosome from each of the subgraphs to be enhanced, and denote it as the second subgraph;

[0034] Select the subgraph to be enhanced located above the diagonal of the chromosome from each of the subgraphs to be enhanced, and denote it as the third subgraph;

[0035] The first subgraph and the third subgraph are input into the generator in the trained generative adversarial network to obtain the enhanced first subgraph and the enhanced third subgraph in the enhanced chromosome conformation contact map output by the generator.

[0036] The matrix formed by the pixel values ​​in the second sub-image is transposed to obtain the transposed second sub-image.

[0037] The transposed second subgraph is input into the trained generator to obtain the enhanced second subgraph output by the generator.

[0038] In one implementation, the step of inputting the contact map of the chromosome conformation to be enhanced into the trained generative adversarial network to obtain the enhanced chromosome conformation contact map output by the trained generative adversarial network further includes:

[0039] Transpose the enhanced second subgraph to obtain the transposed enhanced second subgraph;

[0040] The enhanced first sub-image, the transposed enhanced second sub-image, and the enhanced third sub-image are spliced ​​together according to their positions in the contact map of the chromosome conformation to be enhanced, to obtain the complete contact map of the enhanced chromosome conformation.

[0041] Secondly, embodiments of the present invention also provide a chromosome conformation contact map enhancement device based on generative adversarial networks, wherein the device comprises the following components:

[0042] The preprocessing module is used to process the original chromosome conformation contact map based on the position of each region in the chromosome, and obtain the sample chromosome conformation contact map.

[0043] The training module is used to train a generative adversarial network based on the contact map of the sample chromosome conformation and the base distance of the contact map of the sample chromosome conformation, so as to obtain the trained generative adversarial network.

[0044] A resolution enhancement module is used to input the contact map of the chromosome conformation to be enhanced into the trained generative adversarial network to obtain the enhanced chromosome conformation contact map output by the generator in the trained generative adversarial network.

[0045] Thirdly, embodiments of the present invention also provide a terminal device, wherein the terminal device includes a memory, a processor, and a chromosome conformation contact map enhancement program based on generative adversarial networks stored in the memory and executable on the processor. When the processor executes the chromosome conformation contact map enhancement program based on generative adversarial networks, it implements the steps of the chromosome conformation contact map enhancement method based on generative adversarial networks described above.

[0046] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a chromosome conformation contact map enhancement program based on generative adversarial networks (GANs). When the chromosome conformation contact map enhancement program based on GANs is executed by a processor, it implements the steps of the chromosome conformation contact map enhancement method based on GANs described above.

[0047] Beneficial effects: This invention fully considers the correlation between base distance and chromosome conformation contact map when training the generative adversarial network. During training, base distance is input as conditional information into the generative adversarial network, enabling the network to differentiate chromosome conformation contact maps with different base distances during training. Consequently, the trained generative adversarial network can selectively enhance the resolution of chromosome conformation contact maps based on the base distance of each contact map, thereby improving the quality of contact maps after resolution enhancement for various base distances. Attached Figure Description

[0048] Figure 1 This is an overall flowchart of the present invention;

[0049] Figure 2 This is a contact diagram of chromosome 1 conformation in an embodiment of the present invention;

[0050] Figure 3 This is a structural diagram of the generator in an embodiment of the present invention;

[0051] Figure 4 This is a structural diagram of the discriminator in an embodiment of the present invention;

[0052] Figure 5 This is a flowchart illustrating the calculation of the discriminator loss function in an embodiment of the present invention;

[0053] Figure 6 This is a schematic diagram of the mean square error in an embodiment of the present invention;

[0054] Figure 7 This is a schematic diagram of the peak signal-to-noise ratio in an embodiment of the present invention;

[0055] Figure 8 This is a schematic diagram of structural similarity in an embodiment of the present invention;

[0056] Figure 9 This is a low-resolution contact diagram in an embodiment of the present invention;

[0057] Figure 10 This is a true high-resolution contact diagram in an embodiment of the present invention;

[0058] Figure 11 This is a contact map with enhanced resolution output from the generative adversarial network in this embodiment of the invention.

[0059] Figure 12 This is a comparison chart of the MSE performance of various models under different base distances in the embodiments of the present invention;

[0060] Figure 13 This is a comparison chart of the PSER performance of various models under different base distances in the embodiments of the present invention;

[0061] Figure 14 This is a comparison chart of the SSIM performance of various models under different base distances in the embodiments of the present invention;

[0062] Figure 15 This is a comparison chart of the MAE performance of different models in cross-cell type experiments at different base distances in the embodiments of the present invention;

[0063] Figure 16 This is a comparison chart of the PSNR performance of different models in cross-cell type experiments at different base distances in the embodiments of the present invention;

[0064] Figure 17 This is a comparison of the SSIM performance of different models in cross-cell type experiments at different base distances in the embodiments of the present invention;

[0065] Figure 18 This is a block diagram illustrating the internal structure of a terminal device provided in an embodiment of the present invention. Detailed Implementation

[0066] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0067] Research has revealed that human chromosomes possess a complex, multi-level, three-dimensional spatial structure. Precise analysis of this structure helps elucidate the mechanisms of gene regulation, thereby enabling the prevention and treatment of genetic diseases. High-throughput chromosome conformation capture (Hi-C) sequencing is currently the primary method for obtaining the three-dimensional structure of chromosomes. The common practice is to construct chromosome conformation contact maps based on Hi-C sequencing data, and then further analyze and statistically analyze these maps to determine the presence of chromosomal abnormalities. High-resolution Hi-C sequencing data requires significant human, material, and time investment, but the chromosome conformation contact maps constructed from high-resolution Hi-C sequencing data can accurately reveal the three-dimensional structure of chromosomes (i.e., high-resolution chromosome conformation contact maps can accurately reveal the three-dimensional structure of chromosomes). Low-resolution Hi-C sequencing data offers advantages in speed and cost, but the chromosome conformation contact maps constructed from low-resolution Hi-C sequencing data cannot accurately reveal the three-dimensional structure of chromosomes and may sometimes even lead to erroneous conclusions. However, most available Hi-C data currently have a resolution of 25 to 100 Mb. Therefore, there is a need to develop methods to improve the resolution of Hi-C contact maps (chromosome conformation contact maps). Existing Hi-C contact map enhancement methods include those based on Convolutional Neural Networks (CNNs). CNN-based contact map enhancement methods typically use mean squared error as the loss function, resulting in over-smoothing, blurred edges, and strong distance dependence in the enhanced contact maps. For example, HiCPlus, a HiC contact map enhancement method based on CNNs proposed in 2018, produces highly blurred edges of topological domains and chromatin loops in the enhanced contact map, which is detrimental to downstream bioinformatics analysis. Furthermore, because CNN-based chromosome conformation contact map enhancement methods do not consider base distances, the quality of the enhanced chromosome conformation contact maps is poor.

[0068] To address the aforementioned technical problems, this invention provides a method and apparatus for enhancing chromosome conformation contact maps based on generative adversarial networks (GANs), solving the problem of poor quality chromosome conformation contact maps caused by existing enhancement methods. Specifically, the original chromosome conformation contact map is first preprocessed to obtain a sample chromosome conformation contact map. Then, the sample chromosome conformation contact map and its corresponding base distances are input into the GAN to be trained, thus training the GAN. After training, the chromosome conformation contact map to be enhanced is input into the trained GAN, allowing the GAN to enhance the resolution of the chromosome conformation contact map. This invention fully considers the influence of base distances on chromosome conformation contact maps; therefore, it can improve the quality of chromosome conformation contact maps after resolution enhancement.

[0069] For example, by removing regions containing genes with minimal influence from the original chromosome conformation contact map, the original chromosome conformation contact map can be used as a sample chromosome conformation contact map to train a generative adversarial network (GAN). If we have sample chromosome conformation contact maps A (with base distance a), B (with base distance b), and C (with base distance c), inputting these three maps and their corresponding base distances into the GAN allows for targeted training based on the different base distances. This enables the trained GAN to recognize chromosome conformation contact maps with different base distances, effectively enhancing the resolution of chromosome conformation contact maps corresponding to different base distances, and ultimately improving the image quality after resolution enhancement.

[0070] Exemplary methods

[0071] The chromosome conformation contact map enhancement method based on generative adversarial networks in this embodiment can be applied to terminal devices, which can be terminal products with computing capabilities, such as computers. In this embodiment, as... Figure 1 As shown, the chromosome conformation contact map enhancement method based on generative adversarial networks specifically includes the following steps S100, S200, and S300:

[0072] S100: Based on the position of each region in the chromosome of the original chromosome conformation contact map, process the original chromosome conformation contact map to obtain the sample chromosome conformation contact map.

[0073] In this embodiment, the original chromosome conformation contact map is constructed in the following manner:

[0074] Raw Hi-C sequencing reads are downloaded from the GEO database, and then the raw Hi-C sequencing read files are merged. The merged file is randomly downsampled by 1 / 16 to create a low-resolution sequencing read file. Using the Juicer tool, a high-resolution Hi-C contact map (used as a real high-resolution sample for training the generative adversarial network) and a low-resolution Hi-C contact map (raw chromosome conformation contact map) are constructed from the low-resolution sequencing read file at 10kb resolution, on a chromosome-by-chromosome basis. Since the raw chromosome conformation contact map contains unimportant parts (parts not significant for disease research through chromosomes), the presence of these parts increases the computational cost of training the generative adversarial network. Therefore, removing these parts improves the training speed of the generative adversarial network. Step S100 includes the following steps S101 to S104:

[0075] S101, normalize the pixel values ​​of all the original chromosome conformation contact maps to obtain the normalized original chromosome conformation contact map C. chrk :

[0076] The reason for normalizing all the original chromosome conformation contact maps is that normalization reduces computational effort and increases the speed of training the generative adversarial network. In this embodiment, the high-resolution Hi-C contact map (high-resolution original chromosome conformation contact map) and the low-resolution Hi-C contact map (low-resolution original chromosome conformation contact map) are normalized as follows:

[0077]

[0078] Where, N chrk For B chrk The maximum pixel value among all pixels in this image, i.e. i and j are B chrk The pixel in the i-th row and j-th column of the image.

[0079]

[0080] M chrk This is a contact diagram of the primitive chromosome conformation of the k-th chromosome, n. chrk The number of contact interaction readings for the k-th chromosome is... N is all n chrk The maximum value in, i.e.

[0081] S102, the normalized original chromosome conformation contact map is divided into blocks according to a set size to obtain each sample sub-map used to characterize each region of the original chromosome conformation contact map.

[0082] In this embodiment, the high-resolution and low-resolution Hi-C contact maps are cropped into several non-overlapping sub-maps of size 40×40 (i.e., 400kb×400kb) (each sample sub-map).

[0083] S103, remove the sample sub-images in each of the sample sub-images except those located above the diagonal of the original chromosome conformation contact map and / or remove the sample sub-images located at the centromere of the chromosome, to obtain the pre-selected sample sub-images.

[0084] The diagonal represents the region on the original chromosome conformation contact map where the base distance is 0. The base distance of the subgraphs located on the diagonal is 0, indicating the interaction pairs within the corresponding interval of the subgraph.

[0085] In one embodiment, all sample sub-images excluding the region above the chromosome diagonal and those including the chromosome centromere region are removed from all sample sub-images. The remaining sub-images are those suitable for training the generative adversarial network (the pre-selected sample sub-images). The reason for removing these sample sub-images is that the region above the diagonal is of significant research value, while the chromosome centromere region is not. The specific reasons are as follows:

[0086] by Figure 2 Taking the chromosome conformation contact map corresponding to chromosome 1 as an example, the centromere corresponding to the chromosome conformation contact map has a very small role in gene regulation and its data is difficult to obtain, so this part of the data is not used when constructing the dataset; the Hi-C contact map is sparse, with a large number of dark blocky regions near the diagonal of the contact map, corresponding to microstructures such as topological domains and chromatin loops, which are of great significance for further bioinformatics analysis, while the regions far from the diagonal are usually lighter in color and have weak interactions, and the size of the topological domains is usually less than 1kb; the Hi-C contact map has strict symmetry, and the value of the (x,y) position of the contact map represents the strength of the contact interaction between the x-th segment and the y-th segment of the base sequence, so the values ​​of the (x,y) position and the (y,x) position are completely equal.

[0087] S104, Remove the sample sub-images in each of the pre-selected sample sub-images whose base distance is greater than or equal to a set distance to obtain the sample chromosome conformation contact map.

[0088] Base distance is the number of base pairs separating two interacting sites (i.e., interacting base pairs) on a chromosome. For example, a base distance of 1 Mb between two interacting sites means that the distance between the two sites is 1 M (1 × 10⁻⁶). 6 ) base pairs.

[0089] This embodiment removes sample subgraphs with base distances greater than or equal to 2Mb because portions of contact maps (sample subgraphs) with base distances exceeding 2Mb often lack significant research value.

[0090] Step S100 removes redundant data through steps S101 to S104, effectively reducing computational effort and thus improving the training speed of the generative adversarial network. In one embodiment, data from chromosomes 1 to 17 are used as the training set, data from chromosomes 18 to 22 are used as the test set, and data from chromosome 23 is discarded to eliminate the influence of gender.

[0091] S200, based on the contact map of the sample chromosome conformation and the base distance of the contact map of the sample chromosome conformation, a generative adversarial network is trained to obtain the trained generative adversarial network.

[0092] In this embodiment, when training the generative adversarial network (GAN), the contact map and the corresponding base distances are input into the GAN. This utilizes the crucial information of base distances, enabling the GAN to differentiate between samples with different base distances and to specifically enhance samples based on the local patterns at different base distances. Step S200 includes the following steps S201 to S209:

[0093] S201, the contact map of the sample chromosome conformation and the base distance of the contact map of the sample chromosome conformation are input into the generator to be trained to obtain the enhanced resolution sample map output by the generator to be trained.

[0094] This embodiment includes five types of sample contact maps: sample chromosome conformation contact maps with a base distance of 0Kb, 40Kb, 80Kb, 120Kb, and 160Kb. A one-hot encoding of length 5 is constructed to represent the base distance of the sample. The 5-length one-hot encoding is copied and concatenated to form a 5×40×40 5-channel tensor. This 5×40×40 5-channel tensor is then merged with the sample (sample chromosome conformation contact map) to obtain a 6×40×40 6-channel tensor, which is used as the input to the model. For example, the first channel is used to input 0Kb, the second channel is used to input 40Kb, the third channel is used to input 80Kb, the fourth channel is used to input 120Kb, the fifth channel is used to input 160Kb, and the sixth channel is used to input the sample chromosome conformation contact map. When there is input in the sixth channel and input in the fifth channel, but no input in the other channels, the generative adversarial network knows that the base distance of the input sample chromosome conformation contact map is 160Kb.

[0095] In another embodiment, the one-hot encoding of the base distance is mapped to a single-channel tensor of the same size as the sample through a fully connected layer, and then matrix addition or dot product is performed with the sample to fuse the base distance as conditional information into the sample. Using a fully connected layer method to fuse conditional information makes the model convergence process more stable.

[0096] In one embodiment, the generator is as follows: Figure 3 As shown, the generator consists of five convolutional kernels and five residual modules. Except for the last layer, all convolutional layers in the generator use 3×3 kernels with a stride of 1. After each sample chromosome conformation contact map is input into the generator, it undergoes the following process:

[0097] (1) A convolution with a kernel size of 3×3 and a stride of 1 is transformed into a 64-channel convolution and then activated by the ReLU function;

[0098] (2) Five residual modules, each containing two convolutions and BN normalization and a residual jump connection, with the number of sample channels always remaining at 64;

[0099] (3) A single convolution with a kernel size of 3×3 and a stride of 1, BN normalization, and jump connection across 5 residual modules;

[0100] (4) A convolution with a kernel size of 3×3 and a stride of 1 results in 128 channels;

[0101] (5) A convolution with a kernel size of 3×3 and a stride of 1 results in 256 channels.

[0102] (6) A convolution with a kernel of 1×1 and a stride of 1 is performed, and the number of channels becomes 1. After passing through the Tanh activation function, a high-resolution sample is output.

[0103] By using (1) to (6), the generator takes a low-resolution sample and its base distance information as input and outputs a high-resolution sample (enhanced resolution sample image).

[0104] S202, the enhanced resolution sample image and the true set resolution image output by the generator to be trained are input into the discriminator to be trained to obtain the generation loss function G_LOSS output by the discriminator to be trained. The generation loss function is used to characterize the degree of difference between the enhanced resolution sample image and the true set resolution image.

[0105] The enhanced resolution sample image is generated by enhancing the resolution of the low-resolution sample chromosome conformation contact image, while the true set resolution image (as the standard image) is the high-resolution Hi-C contact image in step S101. The enhanced resolution sample image and the true set resolution image have the same resolution, but the quality of the two images (the degree of image quality, such as the size of the pixel variance) is different. Therefore, a discriminator is needed to determine the degree of quality difference between the two images.

[0106] In one embodiment,

[0107]

[0108] D is the discriminator, c is the base distance, x is the enhanced resolution sample image output by the generator, D(x|c) is the output result of the discriminator D on the input sample x under the condition of the input base distance c, MSE_loss is the mean squared error loss, VGG_loss is the VGG feature loss, TV_loss is the total variational loss, and α is the coefficient of the mean squared error loss, which is an adjustable hyperparameter.

[0109] β is the coefficient of the VGG feature loss, an adjustable hyperparameter. γ is the coefficient of the total variational loss, also an adjustable hyperparameter. In one embodiment, ω = 0.55, α = 0.25, β = 0.05, and γ = 0.15.

[0110] The generated samples are marked as 0, and the mean squared error loss reflects the pixel-level error between the high-resolution samples and the generated samples. The feature layers of the VGG16 network can extract high-level features of the image, such as topological domains and chromatin loops. Therefore, VGG feature loss is introduced into the loss function. Total variation loss is a common method used in the image domain to reduce noise. The Hi-C contact map also has noise caused by technical issues. Therefore, total variation loss is introduced to reduce the impact of noise on the contact map.

[0111] S203, Based on the generation loss function, adjust the parameters of the generator to be trained to complete one round of training of the generator.

[0112] In one embodiment, the parameters of the generator to be trained are adjusted using the optimizer Adam, with an initial learning rate set to 1e-5 (representing 1×10⁻⁵). -5 At 500 rounds, the learning rate decreases to 1e-6, and at 1200 rounds, the learning rate decreases to 5e-7.

[0113] S204, the contact map of the sample chromosome conformation and the base distance of the contact map of the sample chromosome conformation are input into the generator after one round of training to obtain the enhanced resolution sample map output by the generator after one round of training.

[0114] S205, the enhanced resolution sample image output by the generator after one round of training and the real set resolution image are input to the discriminator to be trained to obtain the output value of the discriminator to be trained.

[0115] In one embodiment, the discriminator is a convolutional neural network, such as Figure 4 As shown, the discriminator consists of five convolutional kernels and outputs the loss value through the following process:

[0116] (1) A convolution with a kernel size of 3×3 and a stride of 1 is performed, resulting in 64 channels, and the channel count is reduced to 64. The channel count is then activated by the LeakyReLU function.

[0117] (2) A convolution with a kernel size of 3×3 and a stride of 2 is performed, keeping the number of channels unchanged at 64, and then undergoing LeakyReLU activation and BN normalization.

[0118] (3) A convolution with a kernel size of 3×3 and a stride of 1 is performed, keeping the number of channels unchanged at 64, and then undergoing LeakyReLU activation and BN normalization.

[0119] (4) A convolution with a kernel size of 3×3 and a stride of 2 is performed, keeping the number of channels unchanged at 64, and then undergoing LeakyReLU activation and BN normalization.

[0120] (5) A convolution with a kernel size of 3×3 and a stride of 1 is performed, keeping the number of channels unchanged at 64, and then undergoing LeakyReLU activation and BN normalization.

[0121] (6) A convolution with a kernel size of 3×3 and a stride of 2 is performed, keeping the number of channels unchanged at 64, and then undergoing LeakyReLU activation and BN normalization.

[0122] (7) A convolution with a kernel size of 3×3 and a stride of 1 is performed, resulting in 128 channels, and the channel is normalized by LeakyReLU activation function and BN.

[0123] (8) Flattened into a 1-dimensional tensor with a length of 3200;

[0124] (9) The fully connected layer is mapped to a 1-dimensional tensor of length 512 and then activated by the LeakyReLU function;

[0125] (10) The fully connected layer is mapped to a 1-dimensional tensor of length 1 (the final output of the discriminator).

[0126] The discriminator outputs a score for the high resolution (enhanced resolution sample image output by the generator). The closer the input sample is to the true distribution, the higher the score output by the discriminator.

[0127] S206, Calculate the true loss value of the enhanced resolution sample image output by the generator after one round of training relative to the true set resolution image.

[0128] S207, calculate the loss value of the output value of the discriminator to be trained relative to the output value of the desired discriminator, denoted as the discriminant loss function D_Loss.

[0129] Steps S201 to S203 have completed one round of training for the generator. Afterwards, the generator parameters are kept constant, and training of the discriminator begins. When a sample image is input into the generator, the generator outputs an enhanced resolution sample image. Figure 5 (generating high-resolution samples in, such as) Figure 5 As shown, both the enhanced resolution sample image and the true high-resolution sample are input into the discriminator, which will then have an output value (this output value characterizes the difference between the enhanced resolution sample image and the true high-resolution sample). The loss value D_Loss between the discriminator's output value and the true loss value (the true loss value of the enhanced resolution sample image and the true high-resolution sample) is calculated.

[0130]

[0131] D represents the discriminator, c represents base distance condition information, x represents the sample input to the discriminator (which could be a real high-resolution sample or a generated sample), and P represents the base distance condition information. real (x) represents the distribution of the true high-resolution sample, P fake (x) represents the distribution of generated samples, GP_loss represents the gradient penalty loss, and λ represents the coefficient of the gradient penalty loss, which is an adjustable hyperparameter (λ is 10). The gradient penalty loss GP_Loss is used in conjunction with Wasserstein loss to satisfy the Lipschitz constraint of Wasserstein loss, enabling the model (Generative Adversarial Network) to converge stably.

[0132] S208, Based on the discriminant loss function, adjust the parameters of the discriminator to be trained to obtain one round of training for the discriminator.

[0133] In one embodiment, the parameters of the discriminator are adjusted using the optimizer Adam, with an initial learning rate set to 1e-5, reduced to 1e-6 after 500 rounds, and reduced to 5e-7 after 1200 rounds.

[0134] S209, based on the generator after one round of training and the discriminator after one round of training, continue to train the generator and the discriminator alternately until the number of iterations reaches the set number, and complete the training of the generative adversarial network.

[0135] Step S209 involves repeating steps S201 to S208, alternating between training the generator and the discriminator. Specifically, the generator and discriminator are trained alternately in a 1:1 ratio until 5000 iterations are completed, thus completing the training of both the generator and the discriminator. The purpose of training the discriminator is to enable it to better supervise the training of the generator.

[0136] In another embodiment, the following principle is used to determine whether the training of the generative adversarial network has been completed:

[0137] During training, the discriminator parameters are first fixed, and the generator is minimized by -D(G(x|c)). Then, the generator parameters are fixed again, and the discriminator is maximized by D(x|c). This maximization and minimization process is iterated alternately, and the samples generated by the generator gradually approach the distribution of real high-resolution samples, thus improving the discriminator's ability to distinguish between real and fake images. Using P... r (x) represents the distribution of the true high-resolution sample, P l (x) represents the distribution of the low-resolution sample, P f (x) represents the distribution of the generated samples. The training process can then be represented as a minimization game between the generator and the discriminator, as shown below:

[0138]

[0139] D is the generation loss function output by the discriminator, G is the enhanced resolution sample image output by the generator, c is the base distance conditional information (base distance), x is the sample, which may be a low-resolution sample, a real high-resolution sample, or a generated sample, and P is the generator output. l (x) represents the distribution of the low-resolution sample, P r (x) represents the distribution of the true high-resolution samples (the true set resolution map), P f (x) represents the distribution of generated samples (enhanced resolution sample maps output by the generator), and λ is the coefficient of the gradient penalty loss, an adjustable hyperparameter. This indicates the derivative with respect to x; This indicates that the input sample x follows a distribution P. r When (x) is true, calculate the expected value of D(x|c). When the input sample x follows a distribution, calculate the expected value of -D(G(x|c)). This indicates that the input sample x follows a distribution P. f When (x), for Find the expected value.

[0140] Ideally, the game between the generator and the discriminator will reach a Nash equilibrium (i.e., the above formula reaches its set value when the generator minimizes -D(G(x|c)) and the discriminator maximizes D(x|c), at which point the training of the generative adversarial network is complete). The generator recovers the distribution of the real high-resolution samples, while the discriminator cannot distinguish between genuine and fake samples. Ultimately, a generator network with good sample generation capabilities is obtained. It learns the correspondence between the distribution of low-resolution samples and the distribution of high-resolution samples, and can recover high-resolution samples from low-resolution samples. This generator is then used to enhance the Hi-C contact map.

[0141] S300, the contact map of the chromosome conformation to be enhanced is input into the trained generative adversarial network to obtain the enhanced chromosome conformation contact map output by the trained generative adversarial network.

[0142] The contact map of the chromosome conformation to be enhanced is input into the generator in the trained generative adversarial network, and the generator enhances the resolution of the contact map. Step S300 includes the following steps S301 to S307:

[0143] S301, the contact map of the chromosome conformation to be enhanced is divided into blocks according to the set size to obtain each sub-map to be enhanced.

[0144] In one embodiment, the contact map of the chromosome conformation to be enhanced is cropped into several 40×40 sub-maps to be enhanced.

[0145] S302, select the sub-graph to be enhanced located at the diagonal position of the chromosome from each of the sub-graphs to be enhanced, and denot it as the first sub-graph; select the sub-graph to be enhanced located below the diagonal of the chromosome from each of the sub-graphs to be enhanced, and denot it as the second sub-graph; select the sub-graph to be enhanced located above the diagonal of the chromosome from each of the sub-graphs to be enhanced, and denot it as the third sub-graph.

[0146] S303, the first subgraph and the third subgraph are input into the generator in the trained generative adversarial network to obtain the enhanced first subgraph and the enhanced third subgraph in the enhanced chromosome conformation contact map output by the generator.

[0147] S304, transpose the matrix formed by the pixel values ​​in the second sub-image to obtain the transposed second sub-image.

[0148] S305, the transposed second subgraph is input into the trained generator to obtain the enhanced second subgraph output by the generator.

[0149] S306, the enhanced second subgraph is transposed to obtain the transposed enhanced second subgraph.

[0150] Steps S302 to S306 involve directly using the generator to enhance samples located on or above the diagonal; for samples located below the diagonal, the samples are first transposed, then the generator is used to enhance the transposed samples, and finally the enhanced samples are transposed again to obtain the final result.

[0151] S307, the enhanced first sub-image, the transposed enhanced second sub-image, and the enhanced third sub-image are spliced ​​together according to the positions of the first sub-image, the second sub-image, and the third sub-image in the contact map of the chromosome conformation to be enhanced, to obtain the complete contact map of the enhanced chromosome conformation.

[0152] Step S307 involves stitching the samples together in their original positions to form a complete enhanced Hi-C contact map.

[0153] The following experiments verify that the Generative Adversarial Network (HiC-cGAN) in this embodiment can effectively improve the quality of the contact map after resolution enhancement:

[0154] HiC-cGAN was trained on chromosomes 1 to 17 of cell type GM12878, and the model was tested on chromosomes 18 to 22 of cell type GM12878, yielding the following results: Figure 6 The mean square error (MSE) shown is as follows: Figure 7 The peak signal-to-noise ratio (PSNR) shown is as follows: Figure 8 The structural similarity (SSIM) is shown. From Figure 6 It can be seen that contact maps of chromosome conformations to be enhanced with different base distances have different mean square errors (mean square error is used to characterize image quality) after enhancement by the generative adversarial network. This proves that there is a necessary relationship between base distance and the quality of the enhanced contact map, so the base distance factor needs to be considered when training the generative adversarial network. The same applies to peak signal-to-noise ratio and structural similarity.

[0155] A sub-map with a base distance of 0 kb was selected to visually demonstrate the enhancement effect of HiC-cGAN (the generative adversarial network in this embodiment) on the Hi-C contact map (chromosome conformation contact map). It can be noted that some dark, blocky regions (centromere regions) only exist in areas such as... Figure 10 The actual high-resolution contact map shown does not exist in, for example, Figure 9The low-resolution contact maps shown are either lighter in color or appear to be less distinct. These dark, blocky regions correspond to microscopic chromosomal three-dimensional structures such as topological domains and chromatin loops, which are of great significance to the work of bioinformatics researchers downstream of the task. The generative adversarial network used in this real-time example generates... Figure 11 The enhanced chromosome conformation contact map shown clearly covers the centromere region, therefore the generative adversarial network of this embodiment can significantly improve the quality of the enhanced contact map.

[0156] The HiC-cGAN (the generative adversarial network in this embodiment) of this example was compared with existing GAN, cGAN, WGAN, WGAN-GP, and hicGAN technologies in a comparative experiment (the above model methods were trained on chromosomes 1 to 17 of cell type GM12878, and the model performance was tested on chromosomes 18 to 22 of cell type GM12878), and the results were as follows. Figure 12 The mean square error (MSE) shown is as follows: Figure 13 The peak signal-to-noise ratio (PSNR) shown is as follows: Figure 14 The structural similarity (SSIM) is shown. Since a smaller mean squared error indicates that the generated samples are closer to the high-resolution samples and the quality of the generated samples is better. Figure 12 As can be seen, the generative adversarial network proposed in this embodiment outperforms all existing models at all five base distances, especially when the base distance is small (0kb, 40kb).

[0157] HiC-cGAN also demonstrated excellent performance in cross-cell type experiments. First, HiC-cGAN and four other GAN models were trained on chromosomes 1-17 of the GM12878 cell type. Then, the performance of different models was tested on chromosomes 1-17 and 18-22 of the K562 and NHEK cell types, respectively. The experimental results are as follows: Figure 15 , Figure 16 and Figure 17 As shown, from Figure 15 , Figure 16 and Figure 17 As can be seen, the HiC-cGAN in this embodiment is superior to the other four GAN models.

[0158] In summary, this invention fully considers the correlation between base distance and chromosome conformation contact maps when training the generative adversarial network (GAN). During training, base distance is input as conditional information into the GAN, enabling it to differentiate between chromosome conformation contact maps with different base distances. Consequently, the trained GAN can selectively enhance the resolution of chromosome conformation contact maps based on their specific base distances, thereby improving the quality of contact maps after resolution enhancement for various base distances.

[0159] Furthermore, this invention, based on the symmetry and sparsity of the Hi-C contact map, reduces the amount of data used for training, employing only data above the diagonal of the contact map with a base distance of less than 2kb. The contact map exhibits different local patterns at different base distances; this invention fully considers this crucial information, using base distance as conditional information to construct a conditional generative adversarial network. The loss function is optimized by combining multiple loss mechanisms, including Wasserstein loss, gradient penalty loss, mean squared error loss, VGG feature loss, and total variational loss, enabling the model to converge stably and achieve better enhancement results.

[0160] Exemplary device

[0161] This embodiment also provides a chromosome conformation contact map enhancement device based on generative adversarial networks, the device comprising the following components:

[0162] The preprocessing module is used to process the original chromosome conformation contact map based on the position of each region in the chromosome, and obtain the sample chromosome conformation contact map.

[0163] The training module is used to train a generative adversarial network based on the contact map of the sample chromosome conformation and the base distance of the contact map of the sample chromosome conformation, so as to obtain the trained generative adversarial network.

[0164] A resolution enhancement module is used to input the contact map of the chromosome conformation to be enhanced into the trained generative adversarial network to obtain the enhanced chromosome conformation contact map output by the generator in the trained generative adversarial network.

[0165] Based on the above embodiments, the present invention also provides a terminal device, the principle block diagram of which can be as follows: Figure 18As shown, the terminal device includes a processor, memory, network interface, display screen, and temperature sensor connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a chromosome conformation contact map enhancement method based on generative adversarial networks. The display screen can be an LCD screen or an e-ink screen. The temperature sensor is pre-installed inside the terminal device to detect the operating temperature of the internal components.

[0166] Those skilled in the art will understand that Figure 18 The schematic diagram shown is only a partial structural diagram related to the present invention and does not constitute a limitation on the terminal device to which the present invention is applied. The specific terminal device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0167] In one embodiment, a terminal device is provided, comprising a memory, a processor, and a chromosome conformation contact map enhancement program based on a generative adversarial network (GAN) stored in the memory and executable on the processor. When the processor executes the chromosome conformation contact map enhancement program based on the GAN, it implements the following operation instructions:

[0168] Based on the location of each region in the chromosome in the original chromosome conformation contact map, the original chromosome conformation contact map is processed to obtain the sample chromosome conformation contact map;

[0169] Based on the contact map of the sample chromosome conformation and the base distance of the contact map of the sample chromosome conformation, a generative adversarial network is trained to obtain the trained generative adversarial network.

[0170] The contact map of the chromosome conformation to be enhanced is input into the trained generative adversarial network to obtain the enhanced chromosome conformation contact map output by the trained generative adversarial network.

[0171] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0172] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for enhancing chromosome conformation contact maps based on generative adversarial networks, characterized in that, include: Based on the location of each region in the chromosome in the original chromosome conformation contact map, the original chromosome conformation contact map is processed to obtain the sample chromosome conformation contact map, which includes contact maps with base distances of 0Kb, 40Kb, 80Kb, 120Kb and 160Kb respectively. Based on the contact map of the sample chromosome conformation and the base distance of the contact map of the sample chromosome conformation, the base distance is input into the generative adversarial network to train the generative adversarial network and obtain the trained generative adversarial network. The contact map of the chromosome conformation to be enhanced is input into the generator in the trained generative adversarial network to obtain the output enhanced chromosome conformation contact map. The process of training a generative adversarial network based on the contact map of the sample chromosome conformation and the base distance of the contact map, to obtain the trained generative adversarial network, includes: The contact map of the sample chromosome conformation and the base distance of the contact map of the sample chromosome conformation are input into the generator to be trained to obtain the enhanced resolution sample map output by the generator to be trained. The enhanced resolution sample image and the true set resolution image output by the generator to be trained are input into the discriminator to be trained to obtain the generation loss function output by the discriminator to be trained. The generation loss function is used to characterize the degree of difference between the enhanced resolution sample image and the true set resolution image. Based on the generation loss function, the parameters of the generator to be trained are adjusted to complete one round of training of the generator; The sample chromosome conformation contact map and the base distance of the sample chromosome conformation contact map are input into the generator after one round of training to obtain the enhanced resolution sample map output by the generator after one round of training. The enhanced resolution sample image output by the generator after one round of training and the real set resolution image are input into the discriminator to be trained to obtain the output value of the discriminator to be trained. Calculate the true loss value of the enhanced resolution sample image output by the generator after one round of training, relative to the true set resolution image; Calculate the loss value of the output value of the discriminator to be trained relative to the output value of the desired discriminator, denoted as the discriminant loss function; Based on the discriminant loss function, the parameters of the discriminator to be trained are adjusted to obtain one round of training for the discriminator; Based on the generator and discriminator trained in one round, the generator and discriminator are trained alternately until the set number of iterations is reached, completing the training of the generative adversarial network. The generator and discriminator constitute the generative adversarial network, and the generator's loss function is... : D represents the discriminator, c represents the base distance, and x represents the enhanced resolution sample image output by the generator. Let MSE_loss be the output of the discriminator D on the input sample x given the input base distance c. MSE_loss is the mean squared error loss, VGG_loss is the VGG feature loss, and TV_loss is the total variational loss. This is the coefficient of the mean squared error loss. The coefficients of the VGG feature loss are... The coefficient of the total variational loss; The following formula is used to determine whether the training of the generative adversarial network is complete: D represents the generation loss function output by the discriminator, G represents the enhanced resolution sample image output by the generator, c represents the base distance, and x represents the sample value. The distribution of low-resolution samples. To determine the distribution of the resolution map in a realistic manner. Let λ be the distribution of the enhanced resolution sample map output by the generator, and λ be the coefficient of the gradient penalty loss.

2. The chromosome conformation contact map enhancement method based on generative adversarial networks as described in claim 1, characterized in that, The process of processing the original chromosome conformation contact map based on the location of each region within the chromosome to obtain a sample chromosome conformation contact map includes: The pixel values ​​of the original chromosome conformation contact map are normalized to obtain the normalized original chromosome conformation contact map. The normalized original chromosome conformation contact map is divided into blocks according to a set size to obtain sample sub-maps that characterize each region of the original chromosome conformation contact map; Remove the sample sub-images located below the diagonal in each of the sample sub-images, as well as the sample sub-images located at the centromere of the chromosome, to obtain the pre-selected sample sub-images; Based on the base distances of the pre-selected sample sub-images, the pre-selected sample sub-images are screened to obtain sample chromosome conformation contact maps.

3. The chromosome conformation contact map enhancement method based on generative adversarial networks as described in claim 2, characterized in that, The process of filtering the pre-selected sample sub-images based on the base distances of the pre-selected sample sub-images to obtain sample chromosome conformation contact maps includes: After removing the sample sub-images whose base distance is greater than or equal to a set distance from the pre-selected sample sub-images, a sample chromosome conformation contact map is obtained.

4. The chromosome conformation contact map enhancement method based on generative adversarial networks as described in claim 1, characterized in that... The step of inputting the contact map of the chromosome conformation to be enhanced into the trained generative adversarial network to obtain the enhanced chromosome conformation contact map output by the trained generative adversarial network includes: The contact map of the chromosome conformation to be enhanced is divided into blocks according to a set size to obtain each sub-map to be enhanced; Select the subgraph to be enhanced located at the diagonal position of the chromosome from each of the subgraphs to be enhanced, and denote it as the first subgraph; Select the subgraph to be enhanced located below the diagonal of the chromosome from each of the subgraphs to be enhanced, and denote it as the second subgraph; Select the subgraph to be enhanced located above the diagonal of the chromosome from each of the subgraphs to be enhanced, and denote it as the third subgraph; The first subgraph and the third subgraph are input into the generator in the trained generative adversarial network to obtain the enhanced first subgraph and the enhanced third subgraph in the enhanced chromosome conformation contact map output by the generator. The matrix formed by the pixel values ​​in the second sub-image is transposed to obtain the transposed second sub-image. The transposed second subgraph is input into the trained generator to obtain the enhanced second subgraph output by the generator.

5. The chromosome conformation contact map enhancement method based on generative adversarial networks as described in claim 4, characterized in that, The step of inputting the contact map of the chromosome conformation to be enhanced into the trained generative adversarial network to obtain the enhanced chromosome conformation contact map output by the trained generative adversarial network further includes: Transpose the enhanced second subgraph to obtain the transposed enhanced second subgraph; The enhanced first sub-image, the transposed enhanced second sub-image, and the enhanced third sub-image are spliced ​​together according to their positions in the contact map of the chromosome conformation to be enhanced, to obtain the complete contact map of the enhanced chromosome conformation.

6. A chromosome conformation contact map enhancement device based on generative adversarial networks, characterized in that, The device comprises the following components: The preprocessing module is used to process the original chromosome conformation contact map based on the position of each region in the chromosome to obtain the sample chromosome conformation contact map, which includes contact maps with base distances of 0Kb, 40Kb, 80Kb, 120Kb and 160Kb respectively. The training module is used to train a generative adversarial network based on the contact map of the sample chromosome conformation and the base distance of the contact map of the sample chromosome conformation. The base distance is input into the generative adversarial network to obtain the trained generative adversarial network. The resolution enhancement module is used to input the contact map of the chromosome conformation to be enhanced into the generator in the trained generative adversarial network to obtain the enhanced chromosome conformation contact map output by the trained generative adversarial network. The process of training a generative adversarial network based on the contact map of the sample chromosome conformation and the base distance of the contact map, to obtain the trained generative adversarial network, includes: The contact map of the sample chromosome conformation and the base distance of the contact map of the sample chromosome conformation are input into the generator to be trained to obtain the enhanced resolution sample map output by the generator to be trained. The enhanced resolution sample image and the true set resolution image output by the generator to be trained are input into the discriminator to be trained to obtain the generation loss function output by the discriminator to be trained. The generation loss function is used to characterize the degree of difference between the enhanced resolution sample image and the true set resolution image. Based on the generation loss function, the parameters of the generator to be trained are adjusted to complete one round of training of the generator; The sample chromosome conformation contact map and the base distance of the sample chromosome conformation contact map are input into the generator after one round of training to obtain the enhanced resolution sample map output by the generator after one round of training. The enhanced resolution sample image output by the generator after one round of training and the real set resolution image are input into the discriminator to be trained to obtain the output value of the discriminator to be trained. Calculate the true loss value of the enhanced resolution sample image output by the generator after one round of training, relative to the true set resolution image; Calculate the loss value of the output value of the discriminator to be trained relative to the output value of the desired discriminator, denoted as the discriminant loss function; Based on the discriminant loss function, the parameters of the discriminator to be trained are adjusted to obtain one round of training for the discriminator; Based on the generator and discriminator trained in one round, the generator and discriminator are trained alternately until the set number of iterations is reached, completing the training of the generative adversarial network. The generator and discriminator constitute the generative adversarial network, and the generator's loss function is... : D represents the discriminator, c represents the base distance, and x represents the enhanced resolution sample image output by the generator. Let MSE_loss be the output of the discriminator D on the input sample x given the input base distance c. MSE_loss is the mean squared error loss, VGG_loss is the VGG feature loss, and TV_loss is the total variational loss. This is the coefficient of the mean squared error loss. The coefficients of the VGG feature loss are... The coefficient of the total variational loss; The following formula is used to determine whether the training of the generative adversarial network is complete: D represents the generation loss function output by the discriminator, G represents the enhanced resolution sample image output by the generator, c represents the base distance, and x represents the sample value. The distribution of low-resolution samples. To determine the distribution of the resolution map in a realistic manner. Let λ be the distribution of the enhanced resolution sample map output by the generator, and λ be the coefficient of the gradient penalty loss.

7. A terminal device, characterized in that, The terminal device includes a memory, a processor, and a chromosome conformation contact map enhancement program based on generative adversarial networks (GANs) stored in the memory and executable on the processor. When the processor executes the chromosome conformation contact map enhancement program based on GANs, it implements the steps of the chromosome conformation contact map enhancement method based on GANs as described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a chromosome conformation contact map enhancement program based on generative adversarial networks (GANs). When the GAN-based chromosome conformation contact map enhancement program is executed by a processor, it implements the steps of the chromosome conformation contact map enhancement method based on GANs as described in any one of claims 1-5.

Images