A bacteriophage directed cultivation method and system based on dynamic environmental control

Abstract

The application relates to a bacteriophage directional culture method and system based on dynamic environment control and belongs to the technical field of bacteriophage culture. A domain discriminator is embedded in the three-dimensional convolutional neural network, the single-cell infection state classification model is continuously optimized, and the feature extractor learns the domain-invariant spectral feature representation through domain adversarial training. After the domain discriminator is embedded and the domain adversarial training is performed, the target domain accuracy is improved. The cross-domain generalization capability enables the technical solution of the application to be quickly deployed and popularized among different bacteriophage-host combinations, different culture conditions and different imaging devices, greatly reducing the application threshold and repeated labeling cost of the system. Based on the SAC deep reinforcement learning algorithm training, the optimal control strategy can automatically determine the best time for inducing lysis and trigger the corresponding environment control intervention action according to the infection progress feedback at the single-cell level.

Description

Technical Field

[0001] This invention relates to the field of phage directed culture technology, and in particular to a method and system for phage directed culture based on dynamic environmental control. Background Technology

[0002] Bacteriophages are a class of viruses that parasitize and multiply on bacteria. They exhibit high host specificity, capable of infecting and lysing specific types of bacteria, and are therefore widely used in the treatment of bacterial diseases, food safety control, and environmental microbial remediation. In the clinical application of phage therapy, as well as in industrial settings such as food biosafety and aquaculture disease prevention, large-scale and precise phage culture is the prerequisite and foundation for realizing its application value.

[0003] After infecting host bacteria, bacteriophages face two crucial life cycle decisions: lytic and lysogenic. During the lytic cycle, the bacteriophage replicates massively and lyses the host cell, releasing progeny. During the lysogenic cycle, the bacteriophage genome integrates into the host chromosome, replicating along with the host without producing progeny bacteriophages. This decision is regulated by multiple factors, including host physiological state, concentrations of quorum sensing signaling molecules (such as AI-2 and AHL), nutritional conditions, and the multiple of infection (MOI). Existing research indicates that bacteriophages can regulate their lytic-lysogenic decisions through an information exchange system similar to bacterial quorum sensing. How to accurately monitor the dynamic changes in the bacteriophage infection state during culture and, based on this, achieve intelligent regulation of the lytic-lysogenic transition is one of the core scientific problems in the field of directed bacteriophage culture.

[0004] However, in current culture techniques, this decision-making relies entirely on natural occurrence. Operators cannot perceive changes in infection status at the single-cell level in real time, nor can they actively intervene and optimize the timing and ratio of lysis-lysogen transition during culture. A few studies have attempted to indirectly infer the life cycle decision-making tendencies of bacteriophages by detecting quorum sensing signaling molecules, but a complete technological chain from signaling molecule detection to accurate identification of infection status, and then to active intervention in lysis-lysogen transition, has not yet been established. Although deep learning methods have been attempted to predict the bacteriophage life cycle, they are mainly based on static predictions of nucleotide sequences and cannot achieve dynamic, real-time identification of infection status during culture. Summary of the Invention

[0005] This invention overcomes the shortcomings of the prior art and provides a method and system for directional phage culture based on dynamic environmental control.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The first aspect of this invention provides a method for directional phage culture based on dynamic environmental control, comprising the following steps: Under laboratory conditions, host bacteria were infected with fluorescently labeled bacteriophages, and label data of single-cell infection status and corresponding hyperspectral image data of cells were obtained by fluorescence microscopy to construct a source domain training set. A single-cell infection status classification model is constructed based on a three-dimensional convolutional neural network and a source domain training set. A domain discriminator is embedded in the three-dimensional convolutional neural network to continuously optimize the single-cell infection status classification model. The single-cell infection status classification model was used to analyze the cell infection status of phage culture media within a preset time period, and to perform dynamic tracking and population heterogeneity analysis of single-cell infection. The results of dynamic tracking of single-cell infection and population heterogeneity analysis were statistically analyzed, and the optimal dynamic environmental control strategy was generated.

[0007] Furthermore, in the phage-directed culture method based on dynamic environmental control, under laboratory conditions, fluorescently labeled phages are used to infect host bacteria. Label data of single-cell infection status and corresponding hyperspectral image data of the cells are acquired through fluorescence microscopy to construct a source domain training set, specifically: Under laboratory conditions, host bacteria were infected with fluorescently labeled bacteriophages. The timing, parameters, and order of collection were set, and label data of single-cell infection status and hyperspectral image data of corresponding cells were obtained by fluorescence microscopy. By spatially aligning the label data of the single-cell infection status with the hyperspectral image data of the corresponding cells, it is ensured that the data of the two modalities come from the same micro-trap and the same cells. Based on the far-red fluorescence signal of the host cell, the single bacterial cell in each micro-trap is automatically segmented from each round of images. The regions of the same cell at different time points are aligned by image registration to construct the full time-series tracking trajectory of each single cell. For each cell at each time point, the average fluorescence intensity value of the six fluorescence channels is extracted to form a six-dimensional fluorescence time series for each cell. Based on the time series change pattern of each fluorescence channel, an infection status classification is established. Combining infection status classification and label data, a single-cell hyperspectral data cube is input, and spectral domain enhancement, spatial domain enhancement, hybrid enhancement, and spectral migration enhancement are performed on the single-cell hyperspectral data cube. A source domain training set is constructed based on the enhanced single-cell hyperspectral data cube.

[0008] Furthermore, in the phage-directed culture method based on dynamic environmental control, a single-cell infection status classification model is constructed based on a three-dimensional convolutional neural network and a source domain training set. A domain discriminator is embedded in the three-dimensional convolutional neural network to continuously optimize the single-cell infection status classification model, specifically as follows: A single-cell infection status classification model is constructed based on a three-dimensional convolutional neural network. The single-cell hyperspectral data cube in the source domain training set is used as the model input. The backbone network of the three-dimensional convolutional neural network adopts four layers of three-dimensional convolution kernels to perform convolution operations in both spatial and spectral dimensions. By using convolution operations, a spatial-spectral joint feature map is extracted. The spatial-spectral joint feature map is then globally averaged in both spatial and spectral dimensions to obtain a global feature vector. This vector is then embedded into a domain discriminator after the global average pooling layer of the 3D-CNN backbone network and before the classification head. The input to the domain discriminator is the global feature vector. The classification branch is responsible for predicting the infection status category, and the domain discriminator branch is responsible for predicting whether the feature vector comes from the source domain or the target domain. The domain discriminator outputs the probability of the feature vector coming from the source domain or the target domain. The source domain data and target domain data are divided according to the probability of the two categories from the source domain or from the target domain. The shallow parameters of the 3D-CNN backbone network are fixed, and the entire network is trained together using source domain data and target domain data. The probability distribution of infection status categories is output through the classification head.

[0009] Furthermore, in the phage-directed culture method based on dynamic environmental control, the single-cell infection status classification model is used to analyze the cell infection status of the phage culture medium within a preset time period, specifically as follows: Collect spectral image data of phage culture medium within a preset time period, and input the spectral image data of phage culture medium within the preset time period into the single-cell infection status classification model for infection status analysis; The single-cell infection status classification model is used to output the probability distribution of infection status categories, and the infection status of each region in the phage culture medium is statistically analyzed based on the probability distribution of infection status categories. A heatmap of cell infection status was constructed based on the infection status of each region in the phage culture medium.

[0010] Furthermore, in the phage-directed culture method based on dynamic environmental control, the dynamic tracking of single-cell infection and analysis of population heterogeneity specifically include: By aligning the same micro-trap images at different time points using an image registration algorithm, a complete time trajectory of each single cell from infection to lysis is constructed, and key parameters are extracted from the time trajectory of each single cell. The key parameters include adsorption latency, nucleic acid injection time point, replication and assembly period duration, lysis time point, and total infection cycle duration. Statistical analysis was performed on key parameters of all monitored cells, and the distribution histograms and box plots of each parameter, the correlation heatmap between parameters, and the automatic labeling of abnormal cells were output. The infection status progress at the single-cell level in the phage culture medium was also statistically analyzed.

[0011] Furthermore, in the phage-directed culture method based on dynamic environmental control, the results of single-cell infection dynamic tracking and population heterogeneity analysis are statistically analyzed, and the optimal dynamic environmental control strategy is generated, specifically: The optimal time for inducing lysis is determined based on the progress of infection status at the single-cell level in the phage culture medium, and the progress of infection status at the single-cell level in the phage culture medium is calculated to determine whether it is greater than the preset infection status progress threshold. When the infection progress at the single-cell level in the phage culture medium exceeds a preset infection progress threshold, a corresponding control command is dynamically triggered based on the single-cell infection, and an optimal dynamic environment control strategy is generated based on the corresponding control command. When the infection progress at the single-cell level in the phage culture medium exceeds a preset infection progress threshold, the current culture environment is maintained.

[0012] Furthermore, in the phage-directed culture method based on dynamic environmental control, combined with the dynamic triggering of corresponding control commands by single-cell infection, an optimal dynamic environmental control strategy is generated according to the corresponding control commands, specifically including: The infection state of the bacteriophage at any given time is defined as a vector, and a state space is constructed. Temperature regulation, nutrient feeding rate regulation, inducer addition, stirring rate regulation, and stirring rate regulation are taken as intervention actions. An action space is constructed based on the intervention action, and a reward function is constructed based on the current valence growth rate, solvent ratio, and energy consumption cost. The SAC deep reinforcement learning algorithm is used for pre-training in a digital twin environment. The agent learns the optimal splitting-source conversion control strategy. The entropy regularization coefficient of the SAC deep reinforcement learning algorithm adopts an adaptive adjustment mechanism to maintain a dynamic balance between exploration and exploitation. The optimal pyrolysis-solution source conversion control strategy is found using an intelligent agent, and corresponding control commands are generated. Based on the corresponding control commands, the optimal dynamic environment control strategy is generated.

[0013] A second aspect of the present invention provides a phage directed culture system based on dynamic environment control, including a memory and a processor. The memory includes a phage directed culture method program based on dynamic environment control. When the phage directed culture method program based on dynamic environment control is executed by the processor, it implements the steps of the phage directed culture method based on dynamic environment control as described in any one of the present invention.

[0014] This invention addresses the shortcomings of the prior art and has the following beneficial effects: This invention employs a domain discriminator within a 3D convolutional neural network. Through domain adversarial training, the feature extractor learns domain-invariant spatial spectral feature representations. Compared to 3D-CNN models without domain adaptation mechanisms, the classification accuracy on new systems drops from 94% in the source domain to 62%. After embedding the domain discriminator and performing domain adversarial training, the target domain accuracy improves to 81%. This cross-domain generalization capability enables the technical solution of this invention to be rapidly deployed and promoted across different phage-host combinations, different culture conditions, and different imaging devices, significantly reducing the application threshold and repetitive annotation costs. The optimal control strategy trained based on the SAC deep reinforcement learning algorithm can automatically determine the optimal timing for inducing lysis and trigger corresponding environmental control interventions based on the infection progress feedback at the single-cell level. Experimental results show that after regulation by the intelligent decision-making system, the phage lysis rate is increased, thereby improving potency yield. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained from these drawings without creative effort.

[0016] Figure 1 A flowchart of the phage directed culture method based on dynamic environmental control is shown. Figure 2 A system block diagram of a phage directional culture system based on dynamic environmental control is shown. Detailed Implementation

[0017] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0018] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0019] like Figure 1 As shown, the first aspect of the present invention provides a method for directional phage culture based on dynamic environmental control, comprising the following steps: Under laboratory conditions, host bacteria were infected with fluorescently labeled bacteriophages, and label data of single-cell infection status and corresponding hyperspectral image data of cells were obtained by fluorescence microscopy to construct a source domain training set. A single-cell infection status classification model is constructed based on a three-dimensional convolutional neural network and a source domain training set. A domain discriminator is embedded in the three-dimensional convolutional neural network to continuously optimize the single-cell infection status classification model. The single-cell infection status classification model was used to analyze the cell infection status of phage culture media within a preset time period, and the dynamic tracking of single-cell infection and population heterogeneity analysis were performed. The results of dynamic tracking of single-cell infection and population heterogeneity analysis were statistically analyzed, and the optimal dynamic environmental control strategy was generated.

[0020] It should be noted that in this invention, the domain discriminator in a 3D convolutional neural network learns a domain-invariant spatial spectral feature representation through domain adversarial training. Compared to a 3D-CNN model without a domain adaptation mechanism, the classification accuracy on the new system drops from 94% in the source domain to 62%. After embedding the domain discriminator and performing domain adversarial training, the target domain accuracy increases to 81%. This cross-domain generalization capability enables the technical solution of this invention to be rapidly deployed and promoted across different phage-host combinations, different culture conditions, and different imaging devices, significantly reducing the application threshold and repetitive annotation costs of the system. The optimal control strategy trained based on the SAC deep reinforcement learning algorithm can automatically determine the best time to induce lysis and trigger corresponding environmental control interventions based on the infection progress feedback at the single-cell level. Experimental results show that after regulation by the intelligent decision-making system, the phage lysis ratio is increased, thereby improving potency yield.

[0021] Furthermore, in the phage-directed culture method based on dynamic environmental control, under laboratory conditions, fluorescently labeled phages are used to infect host bacteria. Label data of single-cell infection status and corresponding hyperspectral image data of the cells are acquired through fluorescence microscopy to construct a source domain training set, specifically: Under laboratory conditions, host bacteria were infected with fluorescently labeled bacteriophages. The timing, parameters, and order of collection were set, and label data of single-cell infection status and hyperspectral image data of corresponding cells were obtained by fluorescence microscopy. By spatially aligning the label data of single-cell infection status with the hyperspectral image data of the corresponding cells, it is ensured that the data of the two modalities come from the same micro-trap and the same cells. Based on the far-red fluorescence signal of the host cell, the single bacterial cell in each micro-trap is automatically segmented from each round of images. The regions of the same cell at different time points are aligned by image registration to construct the full time series tracking trajectory of each single cell. For each cell at each time point, the average fluorescence intensity value of the six fluorescence channels is extracted to form a six-dimensional fluorescence time series for each cell. Based on the time series change pattern of each fluorescence channel, an infection status classification is established. Combining infection status classification and label data, input single-cell hyperspectral data cubes, and perform spectral domain enhancement, spatial domain enhancement, hybrid enhancement, and spectral migration enhancement on the single-cell hyperspectral data cubes. Construct source domain training sets based on the enhanced single-cell hyperspectral data cubes.

[0022] It should be noted that the infection status includes uninfected, adsorption complete, nucleic acid injection successful, replication and assembly phase, maturation phase, lysis phase, and lysate establishment phase. Data augmentation is performed using the following methods: Spectral domain enhancement: Random band dropping (randomly dropping 5 to 10 bands and then filling them with linear interpolation) is performed on hyperspectral data to simulate the fluctuation of spectral resolution under different imaging conditions. Spatial domain enhancement: random rotation (0°, 90°, 180°, 270°), random horizontal / vertical flip, random scaling (0.8 to 1.2x), random translation (±2 pixels); MixUp: Linear interpolation is performed on hyperspectral data of cells in different infection states to generate novel training samples at the state transition boundary, thereby improving the model's ability to discriminate boundary conditions; Spectral migration enhancement: Gaussian noise of different levels (randomly selected from SNR=10 to 50dB) is added to hyperspectral data to simulate noise level changes under different imaging conditions.

[0023] It should be noted that by employing spectral domain, spatial domain, and hybrid enhancement strategies, the training set covers a wider range of imaging condition variations, effectively improving the model's generalization ability under different experimental conditions. Furthermore, the constructed source domain training set encompasses various MOI conditions and multiple phage-host combinations, providing a solid data foundation for the subsequent rapid deployment of the model to new culture systems through transfer learning.

[0024] Furthermore, in the phage-directed culture method based on dynamic environmental control, a single-cell infection status classification model is constructed based on a three-dimensional convolutional neural network and a source domain training set. A domain discriminator is embedded in the three-dimensional convolutional neural network to continuously optimize the single-cell infection status classification model. Specifically: A single-cell infection status classification model is constructed based on a three-dimensional convolutional neural network. The single-cell hyperspectral data cube in the source domain training set is used as the model input. The backbone network of the three-dimensional convolutional neural network adopts four layers of three-dimensional convolution kernels to perform convolution operations in both spatial and spectral dimensions. By using convolution operations, a spatial-spectral joint feature map is extracted. The spatial-spectral joint feature map is then globally averaged in both spatial and spectral dimensions to obtain a global feature vector. This vector is then embedded into a domain discriminator after the global average pooling layer of the 3D-CNN backbone network and before the classification head. The input to the domain discriminator is the global feature vector. The classification branch is responsible for predicting the infection status category, and the domain discriminator branch is responsible for predicting whether the feature vector comes from the source domain or the target domain. The domain discriminator outputs the probability of the feature vector coming from the source domain or the target domain. The source domain data and target domain data are divided according to the probability of the two categories from the source domain or from the target domain. The shallow parameters of the 3D-CNN backbone network are fixed, and the entire network is trained together using source domain data and target domain data. The probability distribution of infection status categories is output through the classification head.

[0025] It should be noted that the rich spatial spectral feature knowledge learned in the source domain training set is efficiently transferred to the target domain, and the model is continuously optimized using the constantly accumulating new data during actual deployment, so that the generalization ability of the model continues to improve with the increase of usage time.

[0026] In the early stages of domain adversarial training, the domain discriminator treats all categories equally for domain discrimination. In the later stages of training, a category-conditional domain discriminator is introduced. This discriminator concatenates the category prediction probabilities output by the classification head with the feature vectors and inputs the result into the discriminator. This allows the discriminator to perform domain discrimination within each category, rather than performing global domain discrimination across the entire feature space. This design prevents "over-alignment" during domain adversarial training from blurring category boundaries—that is, it prevents features from different categories from being incorrectly mixed together during domain alignment.

[0027] It should be noted that this invention embeds the domain discriminator after the global average pooling layer and before the classification head, rather than in a shallower or deeper layer. This precise placement ensures that the domain discriminator receives high-level abstract features after sufficient spatial spectral feature extraction, rather than raw pixel-level information or overly abstract category-level information. This guarantees that domain adversarial training can perform domain alignment at the "semantic level" rather than the "pixel level," avoiding the loss of category information caused by over-alignment.

[0028] Furthermore, in the phage-directed culture method based on dynamic environmental control, a single-cell infection status classification model is used to analyze the cell infection status of the phage culture medium within a preset time period, specifically as follows: Collect spectral image data of phage culture medium within a preset time period, and input the spectral image data of phage culture medium within the preset time period into the single-cell infection status classification model for infection status analysis; The probability distribution of infection status categories is output using a single-cell infection status classification model, and the infection status of each region in the phage culture medium is statistically analyzed based on the probability distribution of infection status categories; for example, 80% of the phage growth regions are already in the mature stage.

[0029] A heatmap of cell infection status was constructed based on the infection status of each region in the phage culture medium.

[0030] It should be noted that the cell infection status heatmap can visually reflect the cell infection status in that area.

[0031] Furthermore, in the phage-directed culture method based on dynamic environmental control, the dynamic tracking of single-cell infection and analysis of population heterogeneity specifically include: By aligning the same micro-trap images at different time points using an image registration algorithm, a complete time trajectory of each single cell from infection to lysis is constructed, and key parameters are extracted from the time trajectory of each single cell. Key parameters include adsorption latency, nucleic acid injection time point, replication and assembly period duration, lysis time point, and total infection cycle duration; Statistical analysis was performed on key parameters of all monitored cells, and the distribution histograms and box plots of each parameter, the correlation heatmap between parameters, and the automatic labeling of abnormal cells were output. The infection status progress at the single-cell level in the phage culture medium was also statistically analyzed.

[0032] Furthermore, in the phage-directed culture method based on dynamic environmental control, the results of single-cell infection dynamic tracking and population heterogeneity analysis are statistically analyzed, and the optimal dynamic environmental control strategy is generated, specifically: The optimal time for inducing lysis is determined based on the progress of infection status at the single-cell level in the phage culture medium, and the progress of infection status at the single-cell level in the phage culture medium is calculated to determine whether it is greater than the preset infection status progress threshold. When the infection progress at the single-cell level in the phage culture medium exceeds the preset infection progress threshold, the corresponding control command is dynamically triggered based on the single-cell infection, and the optimal dynamic environment control strategy is generated according to the corresponding control command. When the progress of infection at the single-cell level in the phage culture medium exceeds the preset infection progress threshold, the current culture environment is maintained.

[0033] It should be noted that this method allows for dynamic adjustments based on actual circumstances, thereby increasing potency yield.

[0034] Furthermore, in the phage-directed culture method based on dynamic environmental control, the optimal dynamic environmental control strategy is generated by combining the dynamic triggering of corresponding control commands with single-cell infection, specifically including: The infection state of the bacteriophage at any given time is defined as a vector, and a state space is constructed. Temperature regulation, nutrient feeding rate regulation, inducer addition, stirring rate regulation, and stirring rate regulation are taken as intervention actions. An action space is constructed based on the intervention action, and a reward function is constructed based on the current valence growth rate, the proportion of dissolved sources, and the energy consumption cost. The SAC deep reinforcement learning algorithm is used for pre-training in a digital twin environment. The reward function is: R = α × current valence growth rate + β × (1 - solvent ratio) - γ × energy cost, where α, β, and γ are adjustable weighting coefficients (default α = 0.5, β = 0.3, γ = 0.2). By adjusting the weighting coefficients, different cultivation objectives can be adapted (e.g., "high-yield mode" emphasizes α, "high-purity mode" emphasizes β, and "energy-saving mode" emphasizes γ).

[0035] The agent learns the optimal splitting-source conversion control strategy. The entropy regularization coefficient of the SAC deep reinforcement learning algorithm adopts an adaptive adjustment mechanism to maintain a dynamic balance between exploration and exploitation. The system utilizes an intelligent agent to find the optimal pyrolysis-solution source conversion control strategy and generates corresponding control commands. Based on these control commands, it then generates the optimal dynamic environment control strategy.

[0036] It should be noted that the optimal control strategy trained based on the SAC deep reinforcement learning algorithm can automatically determine the best time to induce lysis and trigger corresponding environmental control interventions based on the infection progress feedback at the single-cell level. Experimental results show that after regulation by the intelligent decision-making system, the phage lysis rate was increased, thereby improving titer yield.

[0037] In addition, this method also includes: In addition to adversarial training, an explicit statistical domain alignment constraint is introduced into the feature space to constrain the distance between the source domain feature mean and the target domain feature mean in the feature space. The maximum mean difference is used as the distance metric in the dimension of the global feature vector. Constrain the norm distance between the source domain feature covariance and the target domain feature covariance, add the statistical alignment loss to the overall loss function, and generate pseudo-labels using the model's prediction results of unlabeled data in the target domain during domain adversarial training. Add high-confidence pseudo-label samples to the training set. Samples of the target domain whose model prediction confidence is greater than a preset confidence threshold are retained and their pseudo-labels are treated as real labels. As the number of training rounds increases, the confidence threshold is gradually reduced so that more samples of the target domain are included in the training. When calculating the classification loss, the loss weight of the pseudo-label sample is its confidence score. In the early stage of domain adversarial training, the domain discriminator performs domain discrimination for all categories equally. In the middle and later stages of training, a category-conditional domain discriminator is introduced. The category prediction probability output by the classification head is concatenated with the feature vector and then input into the domain discriminator, so that the domain discriminator performs domain discrimination within each category.

[0038] It's important to note that the class prediction probabilities output by the classification head are concatenated with the feature vectors and then input into the domain discriminator. This allows the discriminator to perform domain discrimination within each class, rather than global domain discrimination across the entire feature space. This design prevents "over-alignment" during domain adversarial training, which can lead to blurred class boundaries—that is, it prevents features from different classes from being incorrectly mixed together during domain alignment. Furthermore, when calculating the classification loss, the loss weight for pseudo-labeled samples is their confidence score; low-confidence samples contribute less to gradient updates, reducing the negative impact of erroneous pseudo-labels. The progressive asymmetric strategy proposed in this method—first global alignment, then fine-grained class-conditional alignment—better handles situations where the class distributions between the source and target domains differ, achieving domain alignment while protecting the class decision boundaries from being compromised.

[0039] In addition, this method also includes: A triplet network architecture is used to map single-cell hyperspectral data to a low-dimensional embedding space, which allows cells with the same infection state to aggregate in the embedding space and cells with different infection states to separate from each other. For the phage-host pair to be evaluated, hyperspectral data of no less than 1000 single cells were collected at fixed time points after infection. Each cell was mapped to the embedding space through an embedding network, and the proportion of cells classified as lysis phase and maturity phase was calculated and defined as the phage lysis efficiency index. Calculate the phage lysis efficiency index under different MOI gradients, plot the phage lysis efficiency index curve, define the MOI corresponding to the first time the phage lysis efficiency index exceeds the threshold as the lowest effective MOI, and construct new phage-host pairing data. The new phage-host pairing data is incrementally updated to the embedded network through a domain adversarial learning mechanism, and is periodically cross-validated with the results of the traditional plaque method.

[0040] It should be noted that quantitative assessment of phage sensitivity has been achieved at single-cell resolution. The phage lysis efficiency index can effectively distinguish the varying lysis capabilities of different phages targeting the same host bacterium, overcoming the limitation of traditional plaque methods which can only provide qualitative judgments. The information on single-cell heterogeneity gives the phage lysis efficiency index higher information value than traditional population averaging methods, revealing the distribution of infection status among different cells within the same pair, and providing a more comprehensive perspective for evaluating the actual bactericidal efficacy of phages.

[0041] It should be noted that, in the specific implementation process, microfluidic chips or micro-trap array chips are used as carriers for cell culture and observation. Host bacteria are inoculated into the micro-trap array, with each micro-trap measuring approximately 5 to 10 μm in size, ensuring that each micro-trap contains a single bacterial cell. Fluorescently labeled bacteriophages (such as those carrying GFP fluorescent labels) are used to infect the host bacteria at a predetermined MOI (multiple of infection, such as MOI = 0.1, 1, 5, 10, etc.).

[0042] Images were acquired using an automated fluorescence microscope at different time points post-infection (e.g., 0, 5, 10, 15, 20, 25, 30, 40, 50, 60, 90, and 120 minutes post-infection). Images from six fluorescence channels were acquired at each time point: the far-red channel (for host cell segmentation), the green channel (for phage labeling), and four other channels (for acquiring hyperspectral information). Acquisition parameters were set as follows: exposure time 50 to 200 ms, gain 1 to 4, and pixel size 0.1 to 0.5 μm / pixel.

[0043] After image acquisition is complete, the following processing is performed: (1) Cell segmentation and tracking. Based on the far-red fluorescence signal, deep learning segmentation algorithms such as U-Net or Cellpose are used to automatically segment individual bacterial cell regions in each micro-trap from each frame of the image. Through image registration algorithms (such as rigid registration based on mutual information or non-rigid registration based on optical flow), the same cell regions at different time points are aligned to form the full-time sequence tracking trajectory of each single cell.

[0044] (2) Fluorescence time series extraction. For each cell at each time point, the average fluorescence intensity value and standard deviation of the six fluorescence channels were extracted to form a six-dimensional fluorescence time series for each cell. By labeling cells with known infection states (verified through synchronous experiments and endpoint lysis detection), a fluorescence time series pattern library corresponding to each infection state was established.

[0045] (3) Construction of hyperspectral data cube. For each cell region, extract its hyperspectral data cube (spatial dimension H×W×spectral dimension C, where H=16 to 32 pixels, W=16 to 32 pixels, C=30 to 50 spectral channels).

[0046] (4) Data augmentation. The hyperspectral data cube is augmented with spectral domain enhancement, spatial domain enhancement, hybrid enhancement and spectral transfer enhancement as described above, so that the training set covers a wider range of imaging condition variations.

[0047] Through the above steps, a source domain training set covering multiple MOI conditions and multiple phage-host combinations is constructed, with a total sample size of no less than 10,000 single-cell data.

[0048] The structure of the backbone network of a 3D convolutional neural network is as follows: First layer of 3D convolution: kernel size 3×3×3, stride 1, output channels 32; The second layer of 3D convolution: kernel size 3×3×3, stride 1, output channels 64; The third 3D convolutional layer has a kernel size of 3×3×3, a stride of 2 (downsampling), and 128 output channels. Fourth layer 3D convolution: kernel size 3×3×3, stride 1, output channels 256; Global average pooling layer: Performs global average pooling in the spatial dimension (H×W) and spectral dimension (C) to output a 256-dimensional global feature vector; Classification head: Two fully connected layers (256→128→7), outputting the probability distribution of 7 infection status categories.

[0049] The domain discriminator structure is as follows: First layer fully connected: 256 → 128; Second layer fully connected: 128 → 64; The third fully connected layer: 64→2, outputs the binary classification probability of the source domain / target domain.

[0050] The training process employs the following strategies: (1) Pre-training stage: Only source domain data is used to train the classification branch with cross-entropy loss, so that the model initially has the ability to classify infection status.

[0051] (2) Domain adversarial training phase: Simultaneously input source domain data and target domain data (target domain data is unlabeled data), fix the shallow parameters of the backbone network (the first two layers), and perform domain adversarial training through a gradient reversal layer (GRL). The classification loss and domain discrimination loss are jointly optimized so that the feature extractor learns domain-invariant feature representations.

[0052] (3) Pseudo-label self-training stage: Pseudo-labels are generated using the model’s prediction results on the target domain data. Samples with confidence scores greater than the set threshold (initially set to 0.9, and then gradually reduced to 0.6) are added to the training set to continue optimizing the model in a semi-supervised manner.

[0053] (4) Category Conditional Domain Discriminant Stage: In the later stages of training (such as after the training rounds reach 60% of the total rounds), the category probabilities output by the classification head are concatenated with the feature vectors and input into the category conditional domain discriminator to perform fine-grained domain alignment within each category.

[0054] Training parameter settings: Learning rate 1×10 -4 The batch size is 32, the optimizer is Adam, the total number of training rounds is 200, and an early stopping strategy is used to prevent overfitting.

[0055] In actual phage culture, spectral image data of the culture system are acquired in real time and input into a trained classification model for analysis. For each single cell in a micro-trap, the probability distribution of its infection status category is output, and the state corresponding to the highest probability is taken as the current state.

[0056] A multimodal image registration algorithm based on SIFT or depth features is used to align images of the same micro-trap at different time points, constructing the temporal trajectory of each single cell. For each single-cell trajectory, the following key parameters are automatically extracted: Adsorption latency: The time interval from phage addition to the appearance of an adsorption signal on the fluorescent label; Nucleic acid injection time point: the time when the phage nucleic acid signal is detected entering the host cell; Replication and assembly period duration: the time from nucleic acid injection to detection of complete assembly of progeny phages; Lysis time point: the time when the host cell loses its integrity and the progeny phage is released; Total infection cycle duration: the complete duration from adsorption to lysis.

[0057] Statistical analysis was performed on key parameters of all monitored cells (no less than 500 single cells), and histograms and box plots of the distribution of each parameter and Pearson / Spearman correlation heatmaps between parameters were plotted. Abnormal cells that deviated from the normal range by more than 2 standard deviations were automatically marked by clustering algorithms (such as DBSCAN or LOF algorithm).

[0058] After obtaining infection status progress statistics at the single-cell level, the system generates control strategies according to the following process: (1) Quantification of infection progress: The sum of the proportions of cells in the mature and lysis phases in the culture system at the current moment is used as the infection progress index P(t).

[0059] (2) Trigger judgment: When P(t) > preset threshold P_threshold (default is 0.6, i.e. 60% of cells enter the maturation or lysis phase), the control decision is triggered.

[0060] (3) State space construction: The infection state of the culture system at time t is defined as the state vector s(t) = [the proportion of cells in each infection state, the rate of change of infection progress, and the current environmental parameters].

[0061] (4) Definition of action space: Action space A includes temperature regulation (continuously adjustable within the range of 25 to 42°C), nutrient feeding rate (continuously adjustable from 0 to 2 times the reference rate), inducing agent addition (such as mitomycin C or ultraviolet induction, binary action), and stirring rate (continuously adjustable from 50 to 300 rpm).

[0062] (5) SAC algorithm training: In the digital twin environment, the reward function R = 0.5 × valence growth rate + 0.3 × (1 - source ratio) - 0.2 × energy cost is used for pre-training. The entropy regularization coefficient λ of the SAC algorithm is initially set to 0.2, and an adaptive adjustment mechanism is adopted: when the policy entropy is lower than the target entropy (set to -dim(A)), λ is increased, and vice versa, λ is decreased, so as to maintain a dynamic balance between exploration and exploitation.

[0063] (6) Strategy deployment: The trained and converged agent strategy is deployed to the actual culture system. Based on the real-time single-cell infection status feedback, the agent outputs control commands at decision intervals of 5 to 10 minutes to drive the actuator to adjust the corresponding environmental parameters.

[0064] (7) Online updates: New single-cell tracking data are continuously accumulated during the culture process, and the model and strategy are incrementally updated regularly (such as after each batch of culture is completed), so that the system performance can be continuously improved as the usage time increases.

[0065] like Figure 2 As shown, a second aspect of the present invention provides a phage directed culture system based on dynamic environment control, including a memory and a processor. The memory includes a phage directed culture method program based on dynamic environment control. When the phage directed culture method program based on dynamic environment control is executed by the processor, it implements any of the steps of the phage directed culture method based on dynamic environment control.

[0066] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0067] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0068] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0069] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0070] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0071] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for directional phage culture based on dynamic environmental control, characterized in that, Includes the following steps: Under laboratory conditions, host bacteria were infected with fluorescently labeled bacteriophages, and label data of single-cell infection status and corresponding hyperspectral image data of cells were obtained by fluorescence microscopy to construct a source domain training set. A single-cell infection status classification model is constructed based on a three-dimensional convolutional neural network and a source domain training set. A domain discriminator is embedded in the three-dimensional convolutional neural network to continuously optimize the single-cell infection status classification model. The single-cell infection status classification model was used to analyze the cell infection status of phage culture media within a preset time period, and to perform dynamic tracking and population heterogeneity analysis of single-cell infection. The results of dynamic tracking of single-cell infection and population heterogeneity analysis were statistically analyzed, and the optimal dynamic environmental control strategy was generated.

2. The method for directional phage culture based on dynamic environmental control according to claim 1, characterized in that, Under laboratory conditions, host bacteria were infected using fluorescently labeled bacteriophages. Label data of single-cell infection status and corresponding hyperspectral image data of the cells were acquired through fluorescence microscopy to construct a source domain training set, specifically: Under laboratory conditions, host bacteria were infected with fluorescently labeled bacteriophages. The timing, parameters, and order of collection were set, and label data of single-cell infection status and hyperspectral image data of corresponding cells were obtained by fluorescence microscopy. By spatially aligning the label data of the single-cell infection status with the hyperspectral image data of the corresponding cells, it is ensured that the data of the two modalities come from the same micro-trap and the same cells. Based on the far-red fluorescence signal of the host cell, the single bacterial cell in each micro-trap is automatically segmented from each round of images. The regions of the same cell at different time points are aligned by image registration to construct the full time-series tracking trajectory of each single cell. For each cell at each time point, the average fluorescence intensity value of the six fluorescence channels is extracted to form a six-dimensional fluorescence time series for each cell. Based on the time series change pattern of each fluorescence channel, an infection status classification is established. Combining infection status classification and label data, a single-cell hyperspectral data cube is input, and spectral domain enhancement, spatial domain enhancement, hybrid enhancement, and spectral migration enhancement are performed on the single-cell hyperspectral data cube. A source domain training set is constructed based on the enhanced single-cell hyperspectral data cube.

3. The method for directional phage culture based on dynamic environmental control according to claim 1, characterized in that, A single-cell infection status classification model is constructed based on a three-dimensional convolutional neural network and a source domain training set. A domain discriminator is embedded in the three-dimensional convolutional neural network to continuously optimize the single-cell infection status classification model. Specifically: A single-cell infection status classification model is constructed based on a three-dimensional convolutional neural network. The single-cell hyperspectral data cube in the source domain training set is used as the model input. The backbone network of the three-dimensional convolutional neural network adopts four layers of three-dimensional convolution kernels to perform convolution operations in both spatial and spectral dimensions. By using convolution operations, a joint spatial-spectral feature map is extracted. The joint spatial-spectral feature map is then globally averaged in both spatial and spectral dimensions to obtain a global feature vector. This vector is then embedded into a domain discriminator after the global average pooling layer of the 3D-CNN backbone network and before the classification head. The input to the domain discriminator is the global feature vector. The classification branch is responsible for predicting the infection status category, and the domain discriminator branch is responsible for predicting whether the feature vector comes from the source domain or the target domain. The domain discriminator outputs the probability of the feature vector coming from the source domain or the target domain. The source domain data and target domain data are divided according to the probability of the two categories from the source domain or from the target domain. The shallow parameters of the 3D-CNN backbone network are fixed, and the entire network is trained together using source domain data and target domain data. The probability distribution of infection status categories is output through the classification head.

4. The method for directional phage culture based on dynamic environmental control according to claim 1, characterized in that, The single-cell infection status classification model was used to analyze the cell infection status of phage cultures within a preset time period, specifically as follows: Collect spectral image data of phage culture medium within a preset time period, and input the spectral image data of phage culture medium within the preset time period into the single-cell infection status classification model for infection status analysis; The single-cell infection status classification model is used to output the probability distribution of infection status categories, and the infection status of each region in the phage culture medium is statistically analyzed based on the probability distribution of infection status categories. A heatmap of cell infection status was constructed based on the infection status of each region in the phage culture medium.

5. The method for directional phage culture based on dynamic environmental control according to claim 1, characterized in that, Dynamic tracking of single-cell infection and analysis of population heterogeneity specifically include: Image registration algorithms are used to align images of the same micro-trap at different time points to construct a complete time trajectory of each single cell from infection to lysis, and key parameters are extracted from the time trajectory of each single cell. The key parameters include adsorption latency, nucleic acid injection time point, replication and assembly period duration, lysis time point, and total infection cycle duration. Statistical analysis was performed on key parameters of all monitored cells, and the distribution histograms and box plots of each parameter, the correlation heatmap between parameters, and the automatic labeling of abnormal cells were output. The infection status progress at the single-cell level in the phage culture medium was also statistically analyzed.

6. The method for directional phage culture based on dynamic environmental control according to claim 5, characterized in that, The results of dynamic tracking of single-cell infection and population heterogeneity analysis were statistically analyzed, and the optimal dynamic environmental control strategy was generated, specifically as follows: The optimal time for inducing lysis is determined based on the progress of infection status at the single-cell level in the phage culture medium, and the progress of infection status at the single-cell level in the phage culture medium is calculated to determine whether it is greater than the preset infection status progress threshold. When the infection progress at the single-cell level in the phage culture medium exceeds a preset infection progress threshold, a corresponding control command is dynamically triggered based on the single-cell infection, and an optimal dynamic environment control strategy is generated based on the corresponding control command. When the infection progress at the single-cell level in the phage culture medium exceeds a preset infection progress threshold, the current culture environment is maintained.

7. The method for directional phage culture based on dynamic environmental control according to claim 6, characterized in that, By combining the dynamic triggering of corresponding control commands in conjunction with single-cell infection, an optimal dynamic environment control strategy is generated based on the corresponding control commands, specifically including: The infection state of the bacteriophage at any given time is defined as a vector, and a state space is constructed. Temperature regulation, nutrient feeding rate regulation, inducer addition, stirring rate regulation, and stirring rate regulation are taken as intervention actions. An action space is constructed based on the intervention action, and a reward function is constructed based on the current valence growth rate, solvent ratio, and energy consumption cost. The SAC deep reinforcement learning algorithm is used for pre-training in a digital twin environment. The agent learns the optimal splitting-source conversion control strategy. The entropy regularization coefficient of the SAC deep reinforcement learning algorithm adopts an adaptive adjustment mechanism to maintain a dynamic balance between exploration and exploitation. The optimal pyrolysis-solution source conversion control strategy is found using an intelligent agent, and corresponding control commands are generated. Based on the corresponding control commands, the optimal dynamic environment control strategy is generated.

8. A phage-directed culture system based on dynamic environmental control, characterized in that, The device includes a memory and a processor. The memory includes a program for a phage-directed culture method based on dynamic environment control. When the processor executes the program for the phage-directed culture method based on dynamic environment control, it implements the steps of the phage-directed culture method based on dynamic environment control as described in any one of claims 1-7.

Images