Method for assessing antibiotic toxicity on green algae based on fourier transform infrared spectroscopy

Abstract

This invention discloses a method for assessing the toxicity of antibiotics to green algae based on Fourier transform infrared spectroscopy, comprising: constructing a multidimensional experimental dataset containing different antibiotic concentrations and exposure times under controlled conditions, and simultaneously acquiring raw infrared spectra and degradation efficiency data of green algae. The spectral data are preprocessed to extract toxic response features characterizing cell damage and metabolic feedback features reflecting metabolic activity, forming a dual-path spectral feature set. A heterogeneous information network containing toxic effect edges and degradation feedback edges is constructed using antibiotic concentration and green algae state as nodes. A toxicity joint inference model is constructed based on a graph neural network, and the model is trained using the heterogeneous information network to achieve synergistic analysis of toxic effects and degradation potential. The trained model is used to analyze the target water area situation, simulate the purification process, and generate a dynamic river purification planning report including nodes for biomass maintenance and auxiliary intervention measures, providing comprehensive and accurate purification decision support.

Description

Technical Field

[0001] This invention relates to the fields of environmental toxicology and spectral analysis, and in particular to a method for assessing the toxicity of antibiotics to green algae based on Fourier transform infrared spectroscopy. Background Technology

[0002] With the widespread use of antibiotics in medicine, aquaculture, and agriculture, their residues enter the aquatic environment through various pathways, posing a continuous threat to aquatic ecosystems. Green algae, as primary producers in aquatic ecosystems, are important model organisms for assessing the environmental toxicity of pollutants. Currently, the mainstream method for assessing the toxicity of antibiotics to green algae still relies on traditional growth inhibition tests. These methods require observing changes in algal growth over several days by counting cell numbers or measuring biomass, and then calculating toxicity parameters such as half-effect concentration. While this process is classic and reliable, it has inherent limitations such as being time-consuming, cumbersome, and unable to reflect early stress responses at the cellular level in real time, and it is even more difficult to reveal the microscopic biochemical mechanisms of toxic effects.

[0003] In recent years, Fourier transform infrared spectroscopy (FTIR) has been explored for toxicity monitoring due to its ability to rapidly and non-destructively acquire the overall biochemical composition "fingerprint" information of biological samples. However, current applications mostly rely on qualitative or semi-quantitative judgments based on the shift or intensity changes of specific spectral peaks. They lack systematic algorithmic processes to automatically and accurately extract sensitive features that quantify the correlation between toxicity intensity and the data from complex high-dimensional spectral data. More importantly, current research perspectives are largely limited to assessing the one-way toxic effects of pollutants on organisms, failing to link the metabolic feedback of organisms under stress with the potential degradation function of pollutants. This prevents a unified framework from simultaneously answering the two key questions of "how toxic" and "can and how can organisms purify the pollutants?", thus hindering the in-depth application of this technology in guiding aquatic ecological restoration practices.

[0004] Therefore, this invention proposes a method for assessing the toxicity of antibiotics to green algae based on Fourier transform infrared spectroscopy, thereby providing reliable decision support for the bioremediation of water bodies contaminated by antibiotics. Summary of the Invention

[0005] This invention overcomes the shortcomings of the prior art and provides a method for assessing the toxicity of antibiotics to green algae based on Fourier transform infrared spectroscopy. Its important purpose is to provide reliable decision support for the bioremediation of water bodies contaminated by antibiotics.

[0006] To achieve the above objectives, the first aspect of this invention provides a method for evaluating the toxicity of antibiotics to green algae based on Fourier transform infrared spectroscopy, comprising: Under controlled conditions, multiple green algae test groups with different antibiotic concentrations were set up and test samples were collected at different time points. The collected test samples were used to collect raw spectral data and efficacy data to form a multidimensional test dataset. Based on the multidimensional experimental dataset, the original spectral data is extracted and preprocessed. After the preprocessing is completed, toxicity response feature extraction and metabolic feedback feature extraction are performed in parallel to generate a dual-path spectral feature set. Combining the aforementioned multidimensional experimental dataset and dual-path spectral feature set, with antibiotic concentration and green algae state as nodes, a heterogeneous information network is constructed by connecting nodes through two types of directed edges: toxicity edge and degradation feedback edge. A toxicity joint reasoning model is constructed based on a graph neural network. The heterogeneous information network is used to perform representation learning to generate network embedding vectors. The network embedding vectors are used as training samples to perform deep learning and training on the toxicity joint reasoning model. The system acquires water quality data of the target water area, uses a trained toxicity joint inference model to assess toxicity and predict degradation efficiency, and combines the available green algae species to simulate degradation trends and plan purification, generating a dynamic river purification planning report.

[0007] In this scheme, under controlled conditions, multiple green algae test groups with different antibiotic concentrations are set up, and test samples are collected at different time points. The collected test samples are used to acquire raw spectral data and efficacy data, forming a multidimensional test dataset, specifically including: Under standardized constant temperature and light culture conditions, pure green algae in the exponential growth phase were aseptically inoculated into a liquid culture medium containing a logarithmic concentration gradient of antibiotics. Simultaneously, a blank control group without antibiotics and a background elimination control group without algae were set up. When each preset sampling time point is reached, standardized parallel sampling is performed on each independent experimental group. Two equal volumes of homogeneous algal liquid samples are quantitatively transferred using a pipette and defined as Class I and Class II samples. For one type of sample, algal cells are collected by low-speed centrifugation and repeatedly washed with an inorganic salt buffer that matches the osmotic pressure of the culture medium bottom to remove antibiotics and culture medium impurities attached to the cell surface, thus obtaining a type of pure algal cells. The pure algal cells were formed into a uniform and dense algal cell film on an infrared transparent substrate by vacuum filtration or drop addition, and then loaded into the sample chamber of a Fourier transform infrared spectrometer for raw spectral data acquisition. The second type of sample was divided into two sub-samples. The first sub-sample was filtered through a water-based microporous membrane and then subjected to high performance liquid chromatography-tandem mass spectrometry to determine the immediate residual concentration of the target antibiotic and calculate the antibiotic degradation rate. The second sub-sample was subjected to cell counting or biomass measurement to obtain the growth inhibition rate of green algae. The original spectral dataset and performance dataset were obtained through data collection throughout the entire experimental cycle. The original spectral dataset and performance dataset were then associated with the experimental group number, collection time, and experimental conditions to construct a multidimensional experimental dataset.

[0008] In this scheme, the extraction and preprocessing of raw spectral data based on the multidimensional experimental dataset, followed by parallel extraction of toxicity response features and metabolic feedback features after preprocessing to generate a dual-path spectral feature set, specifically includes: A multidimensional experimental dataset is obtained. Raw spectral data is extracted from the multidimensional experimental dataset and a raw spectral matrix is ​​generated. The Savitzky-Golay convolution smoothing algorithm is used to filter each raw spectrum to obtain smoothed spectral data. An adaptive iterative reweighted penalized least squares algorithm is applied to automatically identify and subtract tilted or curved baselines to obtain the baseline-corrected spectrum. The baseline-corrected spectra are transformed using standard normal variables. By centering and scaling each data point of each sample spectrum, the spectral data are standardized to the same scale, resulting in a standardized spectral data matrix. Toxicity response features and metabolic feedback features are then extracted. A nonnegative matrix factorization algorithm is introduced to decompose the standardized spectral data matrix into a feature basis matrix representing the spectra of basic biochemical components and a coefficient matrix representing the relative abundance of each component, thereby extracting the component proportion features characterizing the relative content changes. By calculating the integral area within the wavenumber intervals corresponding to the characteristic peaks of lipids and proteins, the intensity ratio of the bands characterizing the toxic effect is obtained, and then combined with the component ratio characteristics to generate a subset of toxic response characteristics characterizing the degree of toxic damage. The standardized spectral data matrix is ​​subjected to second derivative transformation, and peak detection is performed on the obtained second derivative spectrum. All the identified local extreme points are taken as potential feature peak positions. The narrow spectral bands centered on each peak position are curve fitted, and the peak position, peak height and peak area of ​​each feature peak are extracted to form a metabolic feedback feature subset characterizing metabolic feedback activity. The toxicity response feature subset and the metabolic feedback feature subset are aligned and spliced ​​according to a unified sample identifier to generate a dual-path spectral feature set.

[0009] In this scheme, the key feature is that, by combining the multidimensional experimental dataset and the dual-path spectral feature set, using antibiotic concentration and algal state as nodes, and connecting the nodes through two types of directed edges—toxicity effect edges and degradation feedback edges—to construct a heterogeneous information network, specifically including: A multidimensional experimental dataset and a dual-path spectral feature set are obtained. The unique identifier, initial antibiotic concentration, exposure time, and corresponding growth inhibition rate and antibiotic degradation rate of each experimental unit are extracted from the multidimensional experimental dataset. The toxic response features and metabolic feedback features that match the experimental unit identifier are obtained from the dual-path spectral feature set, and the heterogeneous nodes are defined and their attributes are encoded. For each experimental unit, an antibiotic concentration node and a green algae state node are created. The node attribute encoding of the antibiotic concentration node is the logarithm of the initial antibiotic concentration and the exposure time. The node attribute encoding of the green algae state node is a feature vector formed by splicing the corresponding toxicity response feature subset and metabolic feedback feature subset. After completing the node definition, construct the toxicity edge from the antibiotic concentration node to the corresponding green algae state node, use the attribute vectors of all antibiotic concentration nodes as training features, use the corresponding growth inhibition rate as training labels, and train a support vector regression model. Using the trained support vector regression model, the relative distance from the sample point to the model decision boundary is calculated for each pair of connected antibiotic concentration nodes and green algae state nodes. The predicted growth inhibition rate and confidence score are obtained. The confidence score is normalized and used as the edge weight of the corresponding toxicity effect edge. Define a virtual node representing global degradation efficiency and construct degradation feedback edges between it and the green algae state nodes. Use the metabolic feedback features of all green algae state nodes as input features and the corresponding antibiotic degradation rate as the target value to train a gradient boosting regression tree model. For each green algae state node, the predicted antibiotic degradation rate of each green algae state node is obtained by using the trained gradient boosting regression tree model. The confidence index is generated by calculating the variance of the predicted values ​​of the sample by the model, and after normalization, it is used as the edge weight of the degradation feedback edge connecting the green algae state node and the degradation efficiency virtual node. Based on the defined antibiotic concentration nodes, green algae state nodes, and degradation efficiency virtual nodes, as well as the established toxicity effect edges and degradation feedback edges containing edge weights, nodes are connected and a heterogeneous information network is constructed.

[0010] In this scheme, the key feature is that the step of constructing a toxicity joint inference model based on a graph neural network, performing representation learning on the heterogeneous information network to generate network embedding vectors, and using the network embedding vectors as training samples to perform deep learning and training on the toxicity joint inference model, specifically includes: Obtain a heterogeneous information network, and extract the attribute codes of all antibiotic concentration nodes, green algae state nodes, and degradation efficiency virtual nodes in the heterogeneous information network and map them to the same feature dimension to construct a node feature set; Two weighted adjacency matrices are constructed based on toxicity edges and degradation feedback edges. The element values ​​of each matrix are defined as the normalized weights of the corresponding edges, generating a set of adjacency relationships that characterize the connection strength and directionality of different types of relationships. A toxicity joint inference model is constructed using a graph neural network architecture. The node feature set and the adjacency relationship set are input into the toxicity joint inference model. A classification edge information transmission mechanism is used to perform representation learning on the heterogeneous information network through multi-layer graph convolution operations. In each graph convolutional layer, the feature information of all neighboring nodes of the connection edge corresponding to each node is weighted and aggregated according to the adjacency relationship, and node update operation is performed. By stacking multiple graph convolutional layers, the complex information of multi-level neighbors is iteratively incorporated to capture long-term dependencies and complex interaction patterns in the network, and the final node updated feature representation is generated. The final node updated feature representation of all green algae state nodes is used as the network embedding vector, and then concatenated with the node features of the corresponding antibiotic concentration node to generate an enhanced node representation vector. The toxicity joint inference model is then trained under supervision based on the enhanced node representation vector. The toxicity joint inference model includes a toxicity assessment head and a degradation efficacy head. The toxicity assessment head maps the enhanced node representation vector to the predicted growth inhibition rate through a fully connected neural network. The degradation efficacy prediction head maps the enhanced node representation vector to the predicted antibiotic degradation rate through a fully connected neural network. The true growth inhibition rate and antibiotic degradation rate obtained from the multidimensional experimental dataset are used as supervision labels. All trainable parameters in the toxicity joint inference model are iteratively updated through the backpropagation algorithm and the adaptive moment estimation optimizer until the convergence condition is met and the desired toxicity joint inference model is output.

[0011] This solution is characterized by acquiring water quality data of the target water area, using a trained toxicity joint inference model for toxicity assessment and degradation efficiency prediction, and combining this with available green algae species for degradation status simulation and purification planning to generate a dynamic river purification planning report, specifically including: Acquire water state data of the target water area, including water volume, flow velocity, initial concentration of the target antibiotic, background water quality parameters, and available green algae species and initial biomass information. Extract features from the water state data of the target water area to generate an initial water state feature vector. The initial state feature vector of the water body is imported into the toxicity joint inference model after training. The initial toxicity level and growth inhibition rate are predicted by the toxicity assessment head, and the initial antibiotic degradation rate is predicted by the degradation efficiency prediction head to generate toxicity joint inference information. Based on the toxicity joint inference information, a degradation trend simulation cycle is performed with a fixed time step. In each simulation time step, the theoretical residual concentration of antibiotics in the water is calculated based on the antibiotic degradation rate predicted at the current time step using a first-level degradation kinetic model. The antibiotic concentration and algal state feature vector for the next time step are updated by the theoretical residual concentration to form the input state vector for the next time step. This vector is then input into the toxicity joint inference model to predict the toxicity level and antibiotic degradation rate at the new time point, thereby simulating the dynamic evolution of the purification process over time. During the dynamic simulation cycle, the purification planning decision is executed in parallel. By monitoring the toxicity level, degradation rate and green algae growth inhibition rate of the model output in each simulation time step in real time, it is determined whether purification intervention is required. If so, the corresponding intervention measures are obtained from the preset intervention measure library through similarity search, and the purification intervention decision for the corresponding simulation time step is generated. When the simulated antibiotic concentration drops below the environmental safety standard or reaches the preset maximum simulation duration, the simulation cycle is terminated. Based on the antibiotic concentration decay curve, toxicity risk level change curve, degradation rate change curve, and all triggered intervention decision points and contents at all time steps on the entire simulation timeline, a dynamic river purification planning report is generated and pushed out.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows: The core innovation of this solution lies in upgrading Fourier transform infrared spectroscopy from a one-way toxicity diagnostic tool into an intelligent analysis system capable of simultaneously analyzing the dynamic closed loop of "stress-response-feedback". By constructing a heterogeneous information network incorporating the bidirectional effects of "antibiotics-green algae" and performing end-to-end joint inference based on graph neural networks, it achieves synergistic and accurate prediction of toxic effects and biodegradation potential. Finally, through dynamic simulation, it generates a river purification planning report that includes the timing of biomass maintenance and auxiliary measures, forming a complete technical closed loop from rapid laboratory assessment to on-site engineering decision-making.

[0013] This invention significantly improves assessment efficiency and information dimensionality, transforming traditional biological testing that takes several days into rapid spectral analysis, while providing mechanistic and functional information. Its final output, a dynamic remediation plan, provides executable and optimizable precise decision support for the remediation of antibiotic pollution in water bodies, achieving an application leap from "identifying problems" to "providing solutions." Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples 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 drawings can be obtained according to these drawings without creative effort.

[0015] Figure 1 A flowchart of the first method for evaluating the toxicity of antibiotics to green algae based on Fourier transform infrared spectroscopy, as provided in an embodiment of the present invention; Figure 2 A second method flowchart of a method for evaluating the toxicity of antibiotics to green algae based on Fourier transform infrared spectroscopy, provided in an embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] 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.

[0017] 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.

[0018] Figure 1 A flowchart of the first method for evaluating the toxicity of antibiotics to green algae based on Fourier transform infrared spectroscopy, as provided in an embodiment of the present invention; like Figure 1 As shown, the present invention provides a first method flowchart for evaluating the toxicity of antibiotics to green algae based on Fourier transform infrared spectroscopy, comprising: S102, under controlled conditions, multiple green algae test groups with different antibiotic concentrations were set up and test samples were collected at different time points. The collected test samples were used to collect raw spectral data and efficacy data to form a multidimensional test dataset. S104, Based on the multidimensional experimental dataset, extract the original spectral data and perform preprocessing. After the preprocessing is completed, perform toxicity response feature extraction and metabolic feedback feature extraction in parallel to generate a dual-path spectral feature set. S106, Combining the multidimensional experimental dataset and the dual-path spectral feature set, using antibiotic concentration and green algae state as nodes, a heterogeneous information network is constructed by connecting nodes through two types of directed edges: toxicity edge and degradation feedback edge. S108, Construct a toxicity joint reasoning model based on graph neural network, perform representation learning on the heterogeneous information network to generate network embedding vectors, and use the network embedding vectors as training samples to perform deep learning and training on the toxicity joint reasoning model. S110: Obtain water quality data of the target water area, use the trained toxicity joint inference model to perform toxicity assessment and degradation efficiency prediction, and combine the available green algae species to simulate degradation status and plan purification, generating a dynamic river purification planning report.

[0019] Furthermore, in a preferred embodiment of the present invention, the step of setting multiple green algae test groups with different antibiotic concentrations under controlled conditions and collecting test samples at different time points, and using the collected test samples to collect raw spectral data and efficacy data to form a multidimensional test dataset, specifically includes: Under standardized constant temperature and light culture conditions, pure green algae in the exponential growth phase were aseptically inoculated into a liquid culture medium containing a logarithmic concentration gradient of antibiotics. Simultaneously, a blank control group without antibiotics and a background elimination control group without algae were set up. When each preset sampling time point is reached, standardized parallel sampling is performed on each independent experimental group. Two equal volumes of homogeneous algal liquid samples are quantitatively transferred using a pipette and defined as Class I and Class II samples. For one type of sample, algal cells are collected by low-speed centrifugation and repeatedly washed with an inorganic salt buffer that matches the osmotic pressure of the culture medium bottom to remove antibiotics and culture medium impurities attached to the cell surface, thus obtaining a type of pure algal cells. The pure algal cells were formed into a uniform and dense algal cell film on an infrared transparent substrate by vacuum filtration or drop addition, and then loaded into the sample chamber of a Fourier transform infrared spectrometer for raw spectral data acquisition. The second type of sample was divided into two sub-samples. The first sub-sample was filtered through a water-based microporous membrane and then subjected to high performance liquid chromatography-tandem mass spectrometry to determine the immediate residual concentration of the target antibiotic and calculate the antibiotic degradation rate. The second sub-sample was subjected to cell counting or biomass measurement to obtain the growth inhibition rate of green algae. The original spectral dataset and performance dataset were obtained through data collection throughout the entire experimental cycle. The original spectral dataset and performance dataset were then associated with the experimental group number, collection time, and experimental conditions to construct a multidimensional experimental dataset.

[0020] It should be noted that the construction of the multidimensional experimental dataset is the foundation for all subsequent analyses. Its core lies in obtaining the biochemical response spectrum and functional efficacy data of green algae simultaneously through standardized controlled experiments, and ensuring the precise spatiotemporal correspondence between the two.

[0021] Specifically, the first step is to establish a series of antibiotic exposure groups covering environmentally relevant concentrations under standardized culture conditions. For example, temperature should be controlled within a specific range, and light intensity and photoperiod should be kept constant to ensure all experimental groups are in the same physiological starting state, eliminating the interference of environmental fluctuations on algal growth and stress response. By setting up logarithmic concentration gradients of antibiotic exposure, the dose-response relationship from no effect to observable effect is comprehensively covered, providing a broad input range for model establishment. A blank control group is established to calibrate the normal growth baseline of green algae, while a background elimination control group is used to quantify and subtract the non-biodegradation (such as photolysis or hydrolysis) of antibiotics in the absence of algae, thus ensuring that subsequent calculations purely reflect the biodegradation contribution of green algae.

[0022] In sample collection and processing, low-speed centrifugation was used to gently separate algal cells to avoid cell rupture, while repeated washing with isotonic buffer (usually three times) was performed to remove antibiotic molecules and culture medium salts adsorbed on the cell surface to the greatest extent possible, preventing them from interfering with subsequent Fourier transform infrared spectroscopy measurements. Preparing the washed algal cells into a uniform thin film is crucial for obtaining high-quality spectra. For example, using zinc selenide crystals as a substrate and vacuum filtration to form a monolayer distribution of cells effectively reduces light scattering and enhances the stability and consistency of the infrared signal. For performance data acquisition, high-performance liquid chromatography-tandem mass spectrometry (HPLC-MS / MS) was used to determine antibiotic residue concentrations. Its high selectivity and sensitivity allow for accurate tracking of concentration changes. Simultaneously, biomass was characterized by direct cell counting or measurement of chlorophyll fluorescence intensity. These two methods were cross-validated to ensure the accuracy of growth inhibition rate calculations. Finally, all data were systematically linked and stored using a unified identifier to form a structured database, providing a solid and traceable data foundation for subsequent feature extraction and model construction.

[0023] Furthermore, in a preferred embodiment of the present invention, the step of extracting and preprocessing the original spectral data based on the multidimensional experimental dataset, and then performing toxicity response feature extraction and metabolic feedback feature extraction in parallel after the preprocessing is completed to generate a dual-path spectral feature set, specifically includes: A multidimensional experimental dataset is obtained. Raw spectral data is extracted from the multidimensional experimental dataset and a raw spectral matrix is ​​generated. The Savitzky-Golay convolution smoothing algorithm is used to filter each raw spectrum to obtain smoothed spectral data. An adaptive iterative reweighted penalized least squares algorithm is applied to automatically identify and subtract tilted or curved baselines to obtain the baseline-corrected spectrum. The baseline-corrected spectra are transformed using standard normal variables. By centering and scaling each data point of each sample spectrum, the spectral data are standardized to the same scale, resulting in a standardized spectral data matrix. Toxicity response features and metabolic feedback features are then extracted. A nonnegative matrix factorization algorithm is introduced to decompose the standardized spectral data matrix into a feature basis matrix representing the spectra of basic biochemical components and a coefficient matrix representing the relative abundance of each component, thereby extracting the component proportion features characterizing the relative content changes. By calculating the integral area within the wavenumber intervals corresponding to the characteristic peaks of lipids and proteins, the intensity ratio of the bands characterizing the toxic effect is obtained, and then combined with the component ratio characteristics to generate a subset of toxic response characteristics characterizing the degree of toxic damage. The standardized spectral data matrix is ​​subjected to second derivative transformation, and peak detection is performed on the obtained second derivative spectrum. All the identified local extreme points are taken as potential feature peak positions. The narrow spectral bands centered on each peak position are curve fitted, and the peak position, peak height and peak area of ​​each feature peak are extracted to form a metabolic feedback feature subset characterizing metabolic feedback activity. The toxicity response feature subset and the metabolic feedback feature subset are aligned and spliced ​​according to a unified sample identifier to generate a dual-path spectral feature set.

[0024] It should be noted that this step extracts the quantitative indicators most relevant to toxic damage and metabolic feedback from complex spectral information, thereby generating a dual-path spectral feature set, which provides a foundation for the subsequent construction of a heterogeneous information network capable of analyzing the complex relationship between the two.

[0025] Specifically, firstly, preprocessing the original spectral matrix is ​​fundamental to ensuring data quality. Savitzky-Golay smoothing effectively filters out high-frequency random noise while preserving the true peak shapes of the spectra. An adaptive iterative reweighted penalized least squares algorithm intelligently identifies and subtracts variable baselines caused by sample scattering or instrument drift, successfully correcting for tilted backgrounds due to algal film inhomogeneity. Subsequent standard normal transformation eliminates multiplicative scattering effects caused by differences in physical states between samples, making all sample spectra comparable. After preprocessing, nonnegative matrix factorization decomposes the standardized spectra into component spectra with clear biochemical significance (such as reference spectra corresponding to lipids, proteins, and polysaccharides) and their relative abundance coefficients, thereby extracting component proportion characteristics reflecting intrinsic compositional changes. Simultaneously, by calculating the integral area ratio of specific wavenumber intervals (such as the lipid CH stretching vibration region of approximately 2920 cm⁻¹ and the protein amide I band of approximately 1650 cm⁻¹), a traditional band intensity ratio characteristic sensitive to external toxic stress is obtained. These two types of features are combined to form a subset of toxicity response features, which together characterize the degree of toxic damage from the intrinsic composition to the level of characteristic functional groups. On the other hand, a second-derivative transformation is performed on the standardized spectrum to enhance the separation of overlapping peaks. By performing peak detection and Lorentz fitting on the transformed spectrum, the peak position, peak height, and area of ​​characteristic peaks related to enzyme activity, such as the amide I band peak, are accurately obtained. These parameters are extremely sensitive to small changes in metabolic state, thus forming a subset of metabolic feedback features. Finally, these two sets of features from the same sample are aligned and spliced ​​according to the sample identifier to form a dual-path spectral feature set that preserves both toxic stress and functional feedback information.

[0026] Furthermore, in a preferred embodiment of the present invention, the step of combining the multidimensional experimental dataset and the dual-path spectral feature set, using antibiotic concentration and green algae state as nodes, and connecting the nodes through two types of directed edges—toxicity effect edges and degradation feedback edges—to construct a heterogeneous information network specifically includes: A multidimensional experimental dataset and a dual-path spectral feature set are obtained. The unique identifier, initial antibiotic concentration, exposure time, and corresponding growth inhibition rate and antibiotic degradation rate of each experimental unit are extracted from the multidimensional experimental dataset. The toxic response features and metabolic feedback features that match the experimental unit identifier are obtained from the dual-path spectral feature set, and the heterogeneous nodes are defined and their attributes are encoded. For each experimental unit, an antibiotic concentration node and a green algae state node are created. The node attribute encoding of the antibiotic concentration node is the logarithm of the initial antibiotic concentration and the exposure time. The node attribute encoding of the green algae state node is a feature vector formed by splicing the corresponding toxicity response feature subset and metabolic feedback feature subset. After completing the node definition, construct the toxicity edge from the antibiotic concentration node to the corresponding green algae state node, use the attribute vectors of all antibiotic concentration nodes as training features, use the corresponding growth inhibition rate as training labels, and train a support vector regression model. Using the trained support vector regression model, the relative distance from the sample point to the model decision boundary is calculated for each pair of connected antibiotic concentration nodes and green algae state nodes. The predicted growth inhibition rate and confidence score are obtained. The confidence score is normalized and used as the edge weight of the corresponding toxicity effect edge. Define a virtual node representing global degradation efficiency and construct degradation feedback edges between it and the green algae state nodes. Use the metabolic feedback features of all green algae state nodes as input features and the corresponding antibiotic degradation rate as the target value to train a gradient boosting regression tree model. For each green algae state node, the predicted antibiotic degradation rate of each green algae state node is obtained by using the trained gradient boosting regression tree model. The confidence index is generated by calculating the variance of the predicted values ​​of the sample by the model, and after normalization, it is used as the edge weight of the degradation feedback edge connecting the green algae state node and the degradation efficiency virtual node. Based on the defined antibiotic concentration nodes, green algae state nodes, and degradation efficiency virtual nodes, as well as the established toxicity effect edges and degradation feedback edges containing edge weights, nodes are connected and a heterogeneous information network is constructed.

[0027] It should be noted that the construction of the heterogeneous information network is the core link connecting experimental observation and computational reasoning. Its purpose is to transform discrete, multidimensional data into a topological structure that can represent the dynamic relationship of "stress-state-function".

[0028] Specifically, firstly, based on the unique identifier of each experimental unit, its exposure conditions (initial concentration, time) and efficacy data (growth inhibition rate, degradation rate) are extracted from the multidimensional experimental dataset. Then, its corresponding toxicity response features (such as lipid-protein ratio, component ratio) and metabolic feedback features (such as characteristic peak parameters) are extracted from the dual-path spectral feature set to complete data alignment. Subsequently, for each experimental unit, two types of nodes are created: the first is the antibiotic concentration node, whose node attribute is encoded as the logarithm of the initial antibiotic concentration and the exposure time of that unit; the second is the green algae state node, whose node attribute is encoded as the complete feature vector formed by concatenating the corresponding subsets of toxicity response features and metabolic feedback features.

[0029] After defining the nodes, a toxicity edge is constructed from the antibiotic concentration node to its corresponding algal state node. The construction and weight assignment of this edge are based on a supervised learning model. Specifically, the attribute vectors of all antibiotic concentration nodes are used as training features, and the corresponding growth inhibition rate obtained from the multidimensional experimental dataset is used as the training label to train a support vector regression model. Using the trained support vector regression model, predictions are made for each pair of connected concentration and state nodes to obtain the predicted growth inhibition rate. Then, the relative distance from the sample point to the model decision boundary is calculated, generating a confidence score. This confidence score is normalized to ensure it falls within a preset weight range, and this value is ultimately set as the weight of the toxicity edge. This process encodes both the strength of the toxicity effect of concentration on the state and the reliability of the model prediction in the edge weight. In parallel, a degradation feedback edge is constructed from the algal state node to a global degradation efficiency virtual node. This virtual node serves as a unified convergence point to represent the output of the degradation function. The weights of the degradation feedback edges use a subset of metabolic feedback features from the attribute vectors of all green algae state nodes as input features, and the corresponding antibiotic removal rate obtained from the multidimensional experimental dataset as the target value. A gradient boosting regression tree model is trained to obtain the predicted removal rate. Then, based on the reliability measure of this prediction result, the weights of the degradation feedback edges connecting the green algae state node and the degradation efficacy virtual node are used. These weights quantify the credibility and strength of the physiological state pointing to the degradation function potential.

[0030] Finally, all antibiotic concentration nodes, algal state nodes, and degradation efficiency virtual nodes defined in the above steps, along with the calculated weighted toxicity edges and degradation feedback edges, are integrated into a graph data structure. This data structure explicitly records the unique identifier, type, and attribute vector of each node, as well as the starting node identifier, ending node identifier, and normalized weight value of each directed edge. Thus, a heterogeneous information network containing two types of nodes, two types of directed edges, and their quantized weights is constructed. This network topology explicitly characterizes the stress path of "antibiotic concentration → algal physiological state" and the feedback path of "algal physiological state → degradation function," providing an input graph containing complex node attributes and relational weights for subsequent representation learning based on graph neural networks.

[0031] Furthermore, in a preferred embodiment of the present invention, the step of constructing a toxicity joint inference model based on a graph neural network, performing representation learning on the heterogeneous information network to generate network embedding vectors, and using the network embedding vectors as training samples to perform deep learning and training on the toxicity joint inference model specifically includes: Obtain a heterogeneous information network, and extract the attribute codes of all antibiotic concentration nodes, green algae state nodes, and degradation efficiency virtual nodes in the heterogeneous information network and map them to the same feature dimension to construct a node feature set; Two weighted adjacency matrices are constructed based on toxicity edges and degradation feedback edges. The element values ​​of each matrix are defined as the normalized weights of the corresponding edges, generating a set of adjacency relationships that characterize the connection strength and directionality of different types of relationships. A toxicity joint inference model is constructed using a graph neural network architecture. The node feature set and the adjacency relationship set are input into the toxicity joint inference model. A classification edge information transmission mechanism is used to perform representation learning on the heterogeneous information network through multi-layer graph convolution operations. In each graph convolutional layer, the feature information of all neighboring nodes of the connection edge corresponding to each node is weighted and aggregated according to the adjacency relationship, and node update operation is performed. By stacking multiple graph convolutional layers, the complex information of multi-level neighbors is iteratively incorporated to capture long-term dependencies and complex interaction patterns in the network, and the final node updated feature representation is generated. The final node updated feature representation of all green algae state nodes is used as the network embedding vector, and then concatenated with the node features of the corresponding antibiotic concentration node to generate an enhanced node representation vector. The toxicity joint inference model is then trained under supervision based on the enhanced node representation vector. The toxicity joint inference model includes a toxicity assessment head and a degradation efficacy head. The toxicity assessment head maps the enhanced node representation vector to the predicted growth inhibition rate through a fully connected neural network. The degradation efficacy prediction head maps the enhanced node representation vector to the predicted antibiotic degradation rate through a fully connected neural network. The true growth inhibition rate and antibiotic degradation rate obtained from the multidimensional experimental dataset are used as supervision labels. All trainable parameters in the toxicity joint inference model are iteratively updated through the backpropagation algorithm and the adaptive moment estimation optimizer until the convergence condition is met and the desired toxicity joint inference model is output.

[0032] It should be noted that the construction and training of the toxicity joint reasoning model aims to transform the complex relationships in heterogeneous information networks into a deep learning model capable of joint prediction.

[0033] Specifically, firstly, based on the constructed heterogeneous information network, its node and edge data are extracted to generate the formatted input required for the graph neural network. The node feature set is obtained by integrating the attribute codes of all nodes in the network: for each green algae state node, its feature is directly taken as the feature vector formed by concatenating the toxicity response feature subset and the metabolic feedback feature subset; for each antibiotic concentration node and the degradation efficiency virtual node, its feature is taken as its corresponding attribute code, namely the logarithm of concentration and exposure time. The original features of all nodes are mapped to a unified feature dimension through a learnable linear transformation layer, thereby forming the node feature set. The adjacency relationship set is then constructed based on the toxicity effect edge and the degradation feedback edge, with the element value of each matrix defined as the normalized weight of the corresponding edge, generating an adjacency relationship set representing the connection strength and directionality of different types of relationships.

[0034] Furthermore, a toxicity joint inference model is constructed using a graph neural network architecture. The node feature set and the adjacency relationship set are input into the toxicity joint inference model, and a categorized edge information transfer mechanism is used to perform representation learning on the heterogeneous information network through multi-layer graph convolution operations. In each graph convolution layer, information transfer and aggregation are processed separately according to edge type. For each target node in the heterogeneous information network, based on the weighted adjacency matrix of toxicity edges, the features of all source nodes pointing to the target through such edges are aggregated; simultaneously, based on the weighted adjacency matrix of degradation feedback edges, the features of all sink nodes pointing from the target through such edges are aggregated. Both aggregations use a weighted average algorithm based on edge weights. Next, the aggregated features of the two types of neighbors are concatenated with the features of the target node itself, and a nonlinear activation function (such as ReLU) and layer normalization operation are used to generate the updated feature representation of the target node in the current graph convolution layer. By stacking multiple such layers, the feature representation of each node can iteratively and weightedly fuse the information of its multi-hop neighbors, thereby capturing long-term dependencies and complex interaction patterns in the network. After multi-layer graph convolutional encoding, all green algae state nodes are selected, and the feature representations output from the final graph convolutional layer are extracted as network embedding vectors. To retain direct experimental condition information in subsequent predictions, the network embedding vector of each green algae state node is concatenated again with the original attributes (initial concentration logarithmic value and exposure time) of its corresponding antibiotic concentration node in the heterogeneous information network to form an enhanced node representation vector.

[0035] Finally, the toxicity joint inference model was trained by enhancing node representation vectors. The model contains two parallel task-specific output heads: a toxicity assessment head and a degradation efficacy prediction head, both sharing features learned from the aforementioned graph neural network. The toxicity assessment head maps the enhanced node representation vectors to the predicted growth inhibition rate through a fully connected neural network; the degradation efficacy prediction head maps them to the predicted antibiotic apparent removal rate through a fully connected neural network. The training objective of the model is to minimize a multi-task joint loss function, which is the weighted sum of the mean squared errors between the predicted and actual growth inhibition rates, and between the predicted and actual removal rates. The actual growth inhibition rate and antibiotic apparent removal rate obtained from the multidimensional experimental dataset are used as supervision labels. Through backpropagation and an adaptive moment estimation optimizer, all trainable parameters in the graph neural network, feature mapping network, and task output heads are iteratively updated until the model converges, resulting in a toxicity joint inference model capable of jointly inferring the toxicity effects and degradation functions of green algae.

[0036] Figure 2 A second method flowchart of a method for evaluating the toxicity of antibiotics to green algae based on Fourier transform infrared spectroscopy, provided in an embodiment of the present invention; like Figure 2 As shown, the present invention provides a second method flowchart for evaluating the toxicity of antibiotics to green algae based on Fourier transform infrared spectroscopy, comprising: S202, acquire water state data of the target water area, including water volume, flow velocity, initial concentration of the target antibiotic, background water quality parameters, and available green algae species and initial biomass information of the target water area, and extract features from the water state data of the target water area to generate an initial state feature vector of the water body. S204, The initial state feature vector of the water body is imported into the toxicity joint inference model after training. The initial toxicity level and growth inhibition rate predicted by the toxicity assessment head are used, and the initial antibiotic degradation rate predicted by the degradation efficiency prediction head is used to generate toxicity joint inference information. S206. Based on the toxicity joint reasoning information, a degradation trend simulation cycle is performed with a fixed time step. In each simulation time step, the theoretical residual concentration of antibiotics in the water is calculated based on the antibiotic degradation rate predicted at the current time step using a first-level degradation kinetic model. S208 updates the antibiotic concentration and green algae state feature vector for the next time step by using the theoretical residual concentration, forming the input state vector for the next time step, and inputs it into the toxicity joint inference model to predict the toxicity level and antibiotic degradation rate at the new time point, so as to simulate the dynamic evolution of the purification process over time. S210, during the dynamic simulation cycle, the purification planning decision is executed in parallel. By monitoring the toxicity level, degradation rate and green algae growth inhibition rate of the model output in each simulation time step in real time, it is determined whether purification intervention is required. If so, the corresponding intervention measures are obtained from the preset intervention measure library through similarity search, and the purification intervention decision for the corresponding simulation time step is generated. S212 When the simulated antibiotic concentration drops below the environmental safety standard or reaches the preset maximum simulation duration, the simulation cycle is terminated. Based on the antibiotic concentration decay curve, toxicity risk level change curve, degradation rate change curve, and all triggered intervention decision points and contents at all time steps on the entire simulation time axis, a dynamic river purification planning report is generated and pushed out.

[0037] It should be noted that acquiring water body status data is the practical basis for planning. Hydrodynamic parameters (such as volume and velocity) are used to estimate hydraulic retention time and simulate mixing states, while background water quality parameters (such as pH and dissolved oxygen) may affect algal activity and the chemical form of antibiotics. When constructing the initial state feature vector of the water body, the aforementioned multi-source heterogeneous data (such as concentration units mg / L and volume units m³) need to be normalized and fused with the baseline spectral feature vector of the selected algal species to form a digital representation that comprehensively characterizes the initial conditions of the system. Inputting this initial state vector into a trained toxicity joint inference model yields the initial assessment. The model, through its dual task heads, simultaneously outputs the initial toxicity level and degradation rate prediction. The degradation status simulation is based on first-order reaction kinetics, and its rate constant directly adopts the real-time value predicted by the model, thus dynamically reflecting the impact of changes in algal activity on purification efficiency. At the end of each time step, the antibiotic concentration is updated based on the calculated theoretical residual concentration, and the green algae state feature vector is simulated and adjusted in conjunction with the currently predicted toxicity effect (e.g., a high growth inhibition rate may correspond to a decline in biomass or metabolic characteristics). This forms the new input for the next time step, realizing a closed-loop iterative simulation of the purification process. Decision logic is executed in parallel during the simulation. The toxicity level, degradation rate, and inhibition rate output by the model are monitored in real time and compared with preset safety thresholds or performance degradation thresholds. Once a condition is triggered, the system will perform a similarity matching search and decision based on the current scenario (e.g., "the degradation rate drops to 40% of the initial value and the toxicity level is medium") in a pre-set intervention measure library. Decision content includes, but is not limited to: suggesting the replacement or supplementation of a specific proportion of green algae biomass at a specific simulation time point; suggesting the timing and dosage of auxiliary purification measures such as activated carbon adsorption or photocatalytic degradation and green algae growth catalysts; or adjusting operating parameters such as hydraulic retention time. Finally, when the simulation reaches the termination condition, all time-series data and intervention events are integrated to generate a planning report that includes concentration decay prediction, risk dynamics, and phased operation plans, providing a visualized and executable technical roadmap for project implementation.

[0038] 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.

[0039] 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.

[0040] 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.

[0041] 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.

[0042] 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 described in 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.

[0043] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for assessing the toxicity of an antibiotic on green algae based on Fourier transform infrared spectroscopy, characterized in that, include: Under controlled conditions, multiple green algae test groups with different antibiotic concentrations were set up and test samples were collected at different time points. The collected test samples were used to collect raw spectral data and efficacy data to form a multidimensional test dataset. Based on the multidimensional experimental dataset, the original spectral data is extracted and preprocessed. After the preprocessing is completed, toxicity response feature extraction and metabolic feedback feature extraction are performed in parallel to generate a dual-path spectral feature set. Combining the aforementioned multidimensional experimental dataset and dual-path spectral feature set, with antibiotic concentration and green algae state as nodes, a heterogeneous information network is constructed by connecting nodes through two types of directed edges: toxicity edge and degradation feedback edge. A toxicity joint reasoning model is constructed based on a graph neural network. The heterogeneous information network is used to perform representation learning to generate network embedding vectors. The network embedding vectors are used as training samples to perform deep learning and training on the toxicity joint reasoning model. The system acquires water quality data of the target water area, uses a trained toxicity joint inference model to assess toxicity and predict degradation efficiency, and combines the available green algae species to simulate degradation trends and plan purification, generating a dynamic river purification planning report.

2. The method for assessing antibiotic toxicity on green algae based on Fourier transform infrared spectroscopy according to claim 1, characterized in that, Under controlled conditions, multiple experimental groups of green algae with different antibiotic concentrations were set up, and experimental samples were collected at different time points. Raw spectral data and efficacy data were collected from the collected experimental samples to form a multidimensional experimental dataset, specifically including: Under standardized constant temperature and light culture conditions, pure green algae in the exponential growth phase were aseptically inoculated into a liquid culture medium containing a logarithmic concentration gradient of antibiotics. Simultaneously, a blank control group without antibiotics and a background elimination control group without algae were set up. When each preset sampling time point is reached, standardized parallel sampling is performed on each independent experimental group. Two equal volumes of homogeneous algal liquid samples are quantitatively transferred using a pipette and defined as Class I and Class II samples. For one type of sample, algal cells are collected by low-speed centrifugation and repeatedly washed with an inorganic salt buffer that matches the osmotic pressure of the culture medium bottom to remove antibiotics and culture medium impurities attached to the cell surface, thus obtaining a type of pure algal cells. The pure algal cells were formed into a uniform and dense algal cell film on an infrared transparent substrate by vacuum filtration or drop addition, and then loaded into the sample chamber of a Fourier transform infrared spectrometer for raw spectral data acquisition. The second type of sample was divided into two sub-samples. The first sub-sample was filtered through a water-based microporous membrane and then subjected to high performance liquid chromatography-tandem mass spectrometry to determine the immediate residual concentration of the target antibiotic and calculate the antibiotic degradation rate. The second sub-sample was subjected to cell counting or biomass measurement to obtain the growth inhibition rate of green algae. The original spectral dataset and performance dataset were obtained through data collection throughout the entire experimental cycle. The original spectral dataset and performance dataset were then associated with the experimental group number, collection time, and experimental conditions to construct a multidimensional experimental dataset.

3. The method for assessing antibiotic toxicity on green algae based on Fourier transform infrared spectroscopy according to claim 1, characterized in that, The process involves extracting raw spectral data from the multidimensional experimental dataset and preprocessing it. After preprocessing, toxicity response feature extraction and metabolic feedback feature extraction are performed in parallel to generate a dual-path spectral feature set, specifically including: A multidimensional experimental dataset is obtained. Raw spectral data is extracted from the multidimensional experimental dataset and a raw spectral matrix is ​​generated. The Savitzky-Golay convolution smoothing algorithm is used to filter each raw spectrum to obtain smoothed spectral data. An adaptive iterative reweighted penalized least squares algorithm is applied to automatically identify and subtract tilted or curved baselines to obtain the baseline-corrected spectrum. The baseline-corrected spectra are transformed using standard normal variables. By centering and scaling each data point of each sample spectrum, the spectral data are standardized to the same scale, resulting in a standardized spectral data matrix. Toxicity response features and metabolic feedback features are then extracted. A nonnegative matrix factorization algorithm is introduced to decompose the standardized spectral data matrix into a feature basis matrix representing the spectra of basic biochemical components and a coefficient matrix representing the relative abundance of each component, thereby extracting the component proportion features characterizing the relative content changes. By calculating the integral area within the wavenumber intervals corresponding to the characteristic peaks of lipids and proteins, the intensity ratio of the bands characterizing the toxic effect is obtained, and then combined with the component ratio characteristics to generate a subset of toxic response characteristics characterizing the degree of toxic damage. The standardized spectral data matrix is ​​subjected to second derivative transformation, and peak detection is performed on the obtained second derivative spectrum. All the identified local extreme points are taken as potential feature peak positions. The narrow spectral bands centered on each peak position are curve fitted, and the peak position, peak height and peak area of ​​each feature peak are extracted to form a metabolic feedback feature subset characterizing metabolic feedback activity. The toxicity response feature subset and the metabolic feedback feature subset are aligned and spliced ​​according to a unified sample identifier to generate a dual-path spectral feature set.

4. The method for assessing antibiotic toxicity to green algae based on Fourier transform infrared spectroscopy according to claim 1, characterized in that, The process combines the multidimensional experimental dataset and the dual-path spectral feature set, using antibiotic concentration and algal state as nodes, and constructs a heterogeneous information network by connecting nodes through two types of directed edges: toxicity edges and degradation feedback edges. Specifically, this includes: A multidimensional experimental dataset and a dual-path spectral feature set are obtained. The unique identifier, initial antibiotic concentration, exposure time, and corresponding growth inhibition rate and antibiotic degradation rate of each experimental unit are extracted from the multidimensional experimental dataset. The toxic response features and metabolic feedback features that match the experimental unit identifier are obtained from the dual-path spectral feature set, and the heterogeneous nodes are defined and their attributes are encoded. For each experimental unit, an antibiotic concentration node and a green algae state node are created. The node attribute encoding of the antibiotic concentration node is the logarithm of the initial antibiotic concentration and the exposure time. The node attribute encoding of the green algae state node is a feature vector formed by splicing the corresponding toxicity response feature subset and metabolic feedback feature subset. After completing the node definition, construct the toxicity edge from the antibiotic concentration node to the corresponding green algae state node, use the attribute vectors of all antibiotic concentration nodes as training features, use the corresponding growth inhibition rate as training labels, and train a support vector regression model. Using the trained support vector regression model, the relative distance from the sample point to the model decision boundary is calculated for each pair of connected antibiotic concentration nodes and green algae state nodes. The predicted growth inhibition rate and confidence score are obtained. The confidence score is normalized and used as the edge weight of the corresponding toxicity effect edge. Define a virtual node representing global degradation efficiency and construct degradation feedback edges between it and the green algae state nodes. Use the metabolic feedback features of all green algae state nodes as input features and the corresponding antibiotic degradation rate as the target value to train a gradient boosting regression tree model. For each green algae state node, the predicted antibiotic degradation rate of each green algae state node is obtained by using the trained gradient boosting regression tree model. The confidence index is generated by calculating the variance of the predicted values ​​of the sample by the model, and after normalization, it is used as the edge weight of the degradation feedback edge connecting the green algae state node and the degradation efficiency virtual node. Based on the defined antibiotic concentration nodes, green algae state nodes, and degradation efficiency virtual nodes, as well as the established toxicity effect edges and degradation feedback edges containing edge weights, nodes are connected and a heterogeneous information network is constructed.

5. The method for evaluating the toxicity of antibiotics to green algae based on Fourier transform infrared spectroscopy according to claim 1, characterized in that, The method for constructing a toxicity joint inference model based on a graph neural network involves performing representation learning on the heterogeneous information network to generate network embedding vectors, and using these network embedding vectors as training samples to perform deep learning and training on the toxicity joint inference model. Specifically, this includes: Obtain a heterogeneous information network, and extract the attribute codes of all antibiotic concentration nodes, green algae state nodes, and degradation efficiency virtual nodes in the heterogeneous information network and map them to the same feature dimension to construct a node feature set; Two weighted adjacency matrices are constructed based on toxicity edges and degradation feedback edges. The element values ​​of each matrix are defined as the normalized weights of the corresponding edges, generating a set of adjacency relationships that characterize the connection strength and directionality of different types of relationships. A toxicity joint inference model is constructed using a graph neural network architecture. The node feature set and the adjacency relationship set are input into the toxicity joint inference model. A classification edge information transmission mechanism is used to perform representation learning on the heterogeneous information network through multi-layer graph convolution operations. In each graph convolutional layer, the feature information of all neighboring nodes of the connection edge corresponding to each node is weighted and aggregated according to the adjacency relationship, and node update operation is performed. By stacking multiple graph convolutional layers, the complex information of multi-level neighbors is iteratively incorporated to capture long-term dependencies and complex interaction patterns in the network, and the final node updated feature representation is generated. The final node updated feature representation of all green algae state nodes is used as the network embedding vector, and then concatenated with the node features of the corresponding antibiotic concentration node to generate an enhanced node representation vector. The toxicity joint inference model is then trained under supervision based on the enhanced node representation vector. The toxicity joint inference model includes a toxicity assessment head and a degradation efficacy head. The toxicity assessment head maps the enhanced node representation vector to the predicted growth inhibition rate through a fully connected neural network. The degradation efficacy prediction head maps the enhanced node representation vector to the predicted antibiotic degradation rate through a fully connected neural network. The true growth inhibition rate and antibiotic degradation rate obtained from the multidimensional experimental dataset are used as supervision labels. All trainable parameters in the toxicity joint inference model are iteratively updated through the backpropagation algorithm and the adaptive moment estimation optimizer until the convergence condition is met and the desired toxicity joint inference model is output.

6. The method for evaluating the toxicity of antibiotics to green algae based on Fourier transform infrared spectroscopy according to claim 1, characterized in that, The process involves acquiring water quality data for the target water area, using a trained toxicity joint inference model for toxicity assessment and degradation efficiency prediction, and combining this with available green algae species for degradation status simulation and purification planning to generate a dynamic river purification planning report. Specifically, this includes: Acquire water state data of the target water area, including water volume, flow velocity, initial concentration of the target antibiotic, background water quality parameters, and available green algae species and initial biomass information. Extract features from the water state data of the target water area to generate an initial water state feature vector. The initial state feature vector of the water body is imported into the toxicity joint inference model after training. The initial toxicity level and growth inhibition rate are predicted by the toxicity assessment head, and the initial antibiotic degradation rate is predicted by the degradation efficiency prediction head to generate toxicity joint inference information. Based on the toxicity joint inference information, a degradation trend simulation cycle is performed with a fixed time step. In each simulation time step, the theoretical residual concentration of antibiotics in the water is calculated based on the antibiotic degradation rate predicted at the current time step using a first-level degradation kinetic model. The antibiotic concentration and algal state feature vector for the next time step are updated by the theoretical residual concentration to form the input state vector for the next time step. This vector is then input into the toxicity joint inference model to predict the toxicity level and antibiotic degradation rate at the new time point, thereby simulating the dynamic evolution of the purification process over time. During the dynamic simulation cycle, the purification planning decision is executed in parallel. By monitoring the toxicity level, degradation rate and green algae growth inhibition rate of the model output in each simulation time step in real time, it is determined whether purification intervention is required. If so, the corresponding intervention measures are obtained from the preset intervention measure library through similarity search, and the purification intervention decision for the corresponding simulation time step is generated. When the simulated antibiotic concentration drops below the environmental safety standard or reaches the preset maximum simulation duration, the simulation cycle is terminated. Based on the antibiotic concentration decay curve, toxicity risk level change curve, degradation rate change curve, and all triggered intervention decision points and contents at all time steps on the entire simulation timeline, a dynamic river purification planning report is generated and pushed out.

Images