Eluting behavior prediction method and system, terminal equipment and storage medium

Abstract

The invention belongs to the technical field of data processing, and discloses an elution behavior prediction method and system, terminal equipment and a storage medium, and the method comprises the following steps: collecting a target liquid separated out in a simulated column chromatography elution process for multiple times at equal time intervals; carrying out near infrared spectrum collection on the obtained simulated collection samples in batches and carrying out preprocessing, training a long-short-term memory network combination model by using a sliding time window mechanism and using a plurality of batches of obtained simulated spectrum data sequenced according to time, and predicting a column chromatography elution process by using the trained model. According to the method, the key nodes in the column chromatography elution process are predicted, so that the key nodes in the elution behavior can be efficiently and accurately identified, and an operator can accurately master the elution time of a target component.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for predicting elution behavior, a terminal device, and a storage medium. Background Technology

[0002] Column chromatography is an important separation and purification technique widely used in chemical, biochemical, and traditional Chinese medicine research. Its basic principle is to utilize the different adsorption capacities between the stationary phase (such as silica gel, ion exchange resin, or gel) and the mobile phase (such as organic solvent or buffer solution), causing different components to have different retention times on the stationary phase, thus achieving separation. In modern industrial production, high-purity and high-efficiency separation technologies are crucial for ensuring product quality, improving production efficiency, and reducing costs. With the increasing demand for high-purity compounds in industries such as pharmaceuticals, biotechnology, and food processing, column chromatography has become an indispensable part of industrial production due to its high-efficiency separation characteristics. For example, in the biopharmaceutical industry, biomolecules such as proteins and antibodies require precise separation to ensure efficacy and safety; in the field of traditional Chinese medicine extraction, complex chemical components require accurate separation to ensure consistent efficacy. Column chromatography, with its high-efficiency separation capabilities, wide applicability, scalability, and integration with modern detection and intelligent optimization technologies, has significant application value in industrial production.

[0003] Monitoring and determining the start and end points of elution in industrial column chromatography primarily rely on a combination of manual experience and physicochemical detection techniques. Technicians typically estimate the optimal collection time based on historical data, experimental records, and elution time windows, and confirm this through periodic sampling (TLC / HPLC). Simultaneously, industrial production often employs online monitoring methods such as UV-Vis, refractive index (RI), and conductivity detection (CD) to track the elution of target compounds. High-precision production may also utilize offline HPLC / LC-MS analysis to ensure separation efficiency. This combination of experience and technology enables industrial column chromatography to achieve basic control and quality monitoring of the elution process during production. Among them, TLC (Thin Layer Chromatography) is a separation technique used to separate different components in a mixture; HPLC (High-Performance Liquid Chromatography) is a more precise and powerful separation technique that can separate, identify, and quantify various components in a sample; UV-Vis (Ultraviolet-Visible Spectroscopy) determines the concentration or structural characteristics of a substance by measuring its absorption of ultraviolet or visible light; and RI (Refractive Index Detection) is a technique that detects substances based on changes in their refractive index.

[0004] Industrial column chromatography relies heavily on manual experience for elution monitoring, but this approach has limitations. Experience-based judgment depends on operator skill and historical data, and is susceptible to batch-to-batch variations, leading to misjudgments of elution start and endpoint. Offline detection methods such as HPLC and UPLC in physicochemical analysis are time-consuming and labor-intensive, unable to provide real-time feedback on concentration changes during elution, resulting in delays in determining elution endpoints and start points. UV-Vis is only suitable for substances with UV absorption, RI is easily affected by solvent gradients, and CD is limited to ionic compounds, restricting its applicability.

[0005] Therefore, existing technologies still need to be improved and enhanced. Summary of the Invention

[0006] This application provides a method and system for predicting elution behavior, a terminal device, and a storage medium, aiming to solve the problems of low prediction efficiency and poor accuracy in the prior art when predicting the start and end points of elution behavior.

[0007] In a first aspect, embodiments of this application provide a method for predicting elution behavior, comprising: The target liquid precipitated during the simulated column chromatography elution process was collected multiple times at equal time intervals to obtain multiple batches of simulated samples sorted by time. Near-infrared spectra were collected from all simulated samples in each batch and preprocessed to obtain multiple batches of simulated spectral data sorted by time. The training set of the simulated spectral data is used in batches to train the Long Short-Term Memory Network Combination Model, so that the trained Long Short-Term Memory Network Combination Model learns the mapping relationship between the temporal feature data and the output target in the simulated spectral data. Using the trained long short-term memory network combined model, based on the mapping relationship and the actual spectral data collected during the column chromatography elution process at the target time, the actual predicted concentration of biomacromolecules at the time following the target time during the column chromatography elution process is predicted, so as to determine the key nodes in the column chromatography elution process.

[0008] In some embodiments, the output target includes: simulated predicted concentrations of biomacromolecules; The step of using a sliding time window mechanism to train the Long Short-Term Memory (LSTM) network ensemble model in batches using the training set from the simulated spectral data, so that the trained LSM network ensemble model learns the mapping relationship between the simulated spectral data and the output target, includes: In each batch, the long short-term memory network combination model is used to predict the training set corresponding to the current time window to obtain the simulated predicted concentration of the biomolecule at the next moment after the current time window. The training set corresponding to the next time window selected after moving a preset distance is predicted to obtain the simulated predicted concentration of the biomolecule at the next time window, until the training set at all times is predicted to obtain the simulated predicted concentration sequence. The simulated predicted concentration sequence is trained until the loss function of the long short-term memory network combined model reaches the target threshold, so that the trained long short-term memory network combined model learns the mapping relationship between the time feature data in the simulated spectral data and the simulated predicted concentration of the biomacromolecule. Each time window has the same preset time step length; the time feature data consists of spectral data sorted by time; the biomacromolecules include proteins, saponins, and antibodies; the key nodes include elution start point and elution end point.

[0009] In some embodiments, the long short-term memory network combined model includes: a two-layer long short-term memory network sub-model, a flattened layer, and a fully connected layer connected in sequence; Each layer of the long short-term memory network sub-model includes a target number of long short-term memory networks connected sequentially. Each long short-term memory network in the first layer of the long short-term memory network sub-model is connected to the long short-term memory network at the same position in the second layer of the long short-term memory network sub-model. Each Long Short-Term Memory (LSTM) network contains hidden units for extracting features from the temporal feature data.

[0010] In some embodiments, the preprocessing includes: white-blackboard correction, absorbance conversion, SG smoothing, first derivative processing, and maximum-minimum normalization processing; The process involves acquiring near-infrared spectra of all simulated samples in each batch and preprocessing them to obtain multiple batches of simulated spectral data sorted by time, including: The target spectral data are obtained by collecting spectral data of all the simulated samples in the target wavelength range using a near-infrared spectrometer. The target spectral data is sequentially subjected to white-black correction, absorbance conversion, SG smoothing, first derivative processing, and maximum-minimum normalization processing to obtain the multiple batches of simulated spectral data sorted by time. The time-sorted simulated spectral data refers to simulated spectral data collected at consecutive target times.

[0011] In some embodiments, the method for predicting the elution behavior further includes: The validation set from the simulated spectral data is input into the trained long short-term memory network ensemble model to calculate the evaluation index, and the structure or hyperparameters of the trained long short-term memory network ensemble model are adjusted according to the calculation results. The evaluation indicators include: coefficient of determination, mean absolute error, root mean square error, and relative prediction bias.

[0012] In some embodiments, the step of using a sliding time window mechanism to train the long short-term memory network ensemble model in batches using the training set in the simulated spectral data further includes: After each prediction is completed on the training set at all times in the current batch, the determination coefficient is calculated using the validation set. If the determination coefficient remains unchanged throughout the training process of consecutive target batches, the training process ends; otherwise, the training process continues until the training set completes prediction for all batches at all times. The completion of prediction of the training set at all times in the current batch is considered the end of the training process for the current batch.

[0013] Secondly, embodiments of this application provide a system for predicting elution behavior, comprising: The acquisition module is used to acquire the target liquid precipitated during the simulated column chromatography elution process multiple times at equal time intervals, and obtain multiple batches of simulated acquisition samples sorted by time. The preprocessing module is used to perform near-infrared spectral acquisition and preprocessing on all simulated samples in each batch to obtain multiple batches of simulated spectral data sorted by time. The training module is used to train the Long Short-Term Memory Network Combination Model on the training set in the simulated spectral data in batches using a sliding time window mechanism, so that the trained Long Short-Term Memory Network Combination Model learns the mapping relationship between the temporal feature data in the simulated spectral data and the output target. The prediction module is used to use the trained long short-term memory network combined model to predict the actual predicted concentration of biomacromolecules at the time following the target time in the column chromatography elution process, based on the mapping relationship and the actual spectral data collected at the target time, so as to determine the key nodes in the column chromatography elution process.

[0014] Thirdly, embodiments of this application provide a real-time detection system for the column chromatography elution process, comprising: The system consists of an online data acquisition unit, a near-infrared spectroscopy acquisition module, a back-end processing system, and a user visualization module, connected sequentially. The online acquisition unit is used to collect the target liquid precipitated during the simulated column chromatography elution process at equal time intervals to obtain simulated samples. The near-infrared spectroscopy acquisition module is used to collect spectral data of the target wavelength range in the simulated sample, obtain target spectral data, and generate a near-infrared characteristic curve based on the target spectral data so that the user visualization module can display it. The background processing system is used to preprocess the target spectral data and predict the actual concentration of biomacromolecules at the next time step after the target time based on the preprocessed target spectral data at the target time. It is also used to issue an early warning signal when the key nodes in the column chromatography elution process are determined based on the near-infrared characteristic curve.

[0015] Fourthly, embodiments of this application provide a terminal device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method for predicting elution behavior as described above.

[0016] Fifthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method for predicting elution behavior as described above.

[0017] Compared to existing technologies, this application provides a method, system, terminal device, and storage medium for predicting elution behavior. The method involves repeatedly collecting the target liquid precipitated during simulated column chromatography elution at equal time intervals, performing near-infrared spectral acquisition on the obtained simulated samples in batches and preprocessing them. A sliding time window mechanism is used to train a long short-term memory network combined model with the multiple batches of time-ordered simulated spectral data. The trained model is then used to predict the column chromatography elution process, identifying key nodes in the elution process. This allows for efficient and accurate identification of key nodes in the elution behavior, enabling operators to grasp the elution timing of the target component and effectively improving the adaptability and production efficiency of the elution process. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart of a method for predicting elution behavior provided in this application; Figure 2 A flowchart illustrating the method for predicting elution behavior provided in this application for obtaining simulated spectral data; Figure 3 A flowchart for predicting the total saponin concentration of American ginseng in the method for predicting elution behavior provided in this application; Figure 4 A structural diagram of the combined long short-term memory network model provided in this application; Figure 5 A structural diagram of an LSTM hidden cell provided in this application; Figure 6 A flowchart illustrating the training of a combined long short-term memory network model in the method for predicting washout behavior provided in this application; Figure 7 A flowchart of an early cessation mechanism in the method for predicting elution behavior provided in this application; Figure 8 A schematic diagram of the structure of the elution behavior prediction system provided in this application; Figure 9 A schematic diagram of a real-time monitoring system for the column chromatography elution process provided in this application; Figure 10 A comparative graph showing the trend of target component concentration changing over time during column chromatography elution provided in this application.

[0020] Figure labels: 10-Acquisition module; 20-Preprocessing module; 30-Training module; 40-Prediction module. Detailed Implementation

[0021] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0022] The components of the embodiments of this application described and illustrated in the accompanying drawings can be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of this application provided in the drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0023] In the following text, the terms "comprising," "having," and their cognates, which may be used in various embodiments of this application, are intended only to indicate a particular feature, number, step, operation, element, component, or combination thereof, and should not be construed as primarily excluding the presence of one or more other features, numbers, steps, operations, elements, components, or combinations thereof, or adding the possibility of one or more combinations thereof. Furthermore, the terms "first," "second," "third," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance.

[0024] Unless otherwise specified, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of this application pertain. Terms (such as those defined in commonly used dictionaries) shall be interpreted as having the same meaning as in their contextual meaning in the relevant technical field and shall not be construed as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of this application.

[0025] This application provides a method, system, terminal device, and storage medium for predicting elution behavior. The method involves repeatedly collecting simulated target liquid precipitated during column chromatography elution at equal time intervals. The obtained simulated samples are then subjected to near-infrared spectral analysis in batches and preprocessed. A sliding time window mechanism is used to train a long short-term memory network (LSTM) combined model with the multiple batches of time-ordered simulated spectral data. The trained model is then used to predict the column chromatography elution process, identifying key nodes in the elution process. This allows for efficient and accurate identification of key nodes in the elution behavior, enabling operators to determine the optimal elution timing for the target component and effectively improving the adaptability and production efficiency of the elution process.

[0026] The design scheme of the prediction method for elution behavior is described below through some specific examples.

[0027] Please see Figure 1 This application provides a method for predicting elution behavior, including steps S100-S400: S100, the target liquid precipitated during the simulated column chromatography elution process is collected multiple times at equal time intervals to obtain multiple batches of simulated samples sorted by time.

[0028] As an example, the process of eluting ginsenosides from ginseng by column chromatography is used: To predict the appropriate starting and ending points of ginseng's column chromatography elution behavior in actual processes, it is necessary to simulate the ginseng column chromatography elution process to train an ideal large-scale model for predicting the actual elution behavior of ginseng in column chromatography. Specifically, the column chromatography elution behavior of total ginsenosides refers to the process in a system using a solid-phase chromatography column as the separation medium, where a sample containing total ginsenosides is loaded, and under the action of the mobile phase (such as aqueous solutions of ethanol at different concentrations), the target components gradually desorb from the stationary phase and elute from the column with the mobile phase.

[0029] The process of training an ideal large model includes: sample acquisition, near-infrared spectral acquisition and preprocessing, and model training.

[0030] Step 1, Sample Collection: During sample collection, the target liquid precipitated during simulated column chromatography elution is collected simultaneously multiple times at equal time intervals (e.g., every 2 minutes), resulting in simulated samples ordered by time, i.e., simulated samples collected at consecutive target times. For example, in simulated ginseng column chromatography elution, a sample (i.e., the target liquid) is collected on average every 2 minutes, with approximately 2 ml collected each time and placed in a centrifuge tube. A total of 25 samples collected, ordered by time, constitute one batch. To avoid randomness and ensure objectivity, eight simultaneous collections are performed each time, resulting in a total of eight batches of samples, i.e., simulated samples. Within each batch, the first six batches serve as the training set, the seventh batch as the validation set, and the eighth batch as the test set.

[0031] It is understood that in this application, the target liquid precipitated during the simulated column chromatography elution process is collected at equal time intervals to obtain simulated collection samples sorted by time, thereby providing training data with dynamic temporal characteristics so that the model can learn dynamic temporal characteristics.

[0032] S200: Near-infrared spectra are collected and preprocessed for all simulated samples in each batch to obtain multiple batches of simulated spectral data sorted by time.

[0033] The preprocessing includes: white-blackboard correction, absorbance conversion, SG smoothing, first derivative processing, and maximum-minimum normalization.

[0034] As an example, after obtaining the simulated sample, the second step is performed: near-infrared spectral acquisition and preprocessing. In each batch, near-infrared spectra of all simulated samples are acquired using a near-infrared spectrometer, and the acquired near-infrared data are preprocessed (including white-black correction, absorbance conversion, SG smoothing, first derivative processing, and maximum-minimum normalization processing, etc.) to obtain simulated spectral data sorted by the same time.

[0035] It is understood that this application uses near-infrared data collected from samples as training data to enable non-contact real-time monitoring of the actual eluent in column chromatography using near-infrared spectroscopy when predicting the start and end points. By collecting spectral data online through a fiber optic probe, dynamic and non-destructive detection of the elution process is achieved, thus preserving the active structure of the biological sample inside the chromatography column.

[0036] In one implementation method, please refer to Figure 2 Step S200: Near-infrared spectral data are acquired and preprocessed for all simulated samples in each batch to obtain multiple batches of simulated spectral data sorted by time, including: S210. Use a near-infrared spectrometer to collect spectral data of the target wavelength range in all simulated samples to obtain target spectral data; S220. The target spectral data is sequentially subjected to white-black correction, absorbance conversion, SG smoothing, first derivative processing, and maximum-minimum normalization processing to obtain multiple batches of simulated spectral data sorted by time.

[0037] Among them, the simulated spectral data sorted by time are simulated spectral data collected at consecutive target times.

[0038] See also, as an example Figure 3 The specific process of near-infrared spectral acquisition and preprocessing is as follows: First, near-infrared spectrometers (e.g., SW2960 type) are used to acquire spectral data of the target wavelength range (e.g., 900–1700 nm) in all simulated samples to obtain the target spectral data. During actual acquisition: the fiber optic probe is fully immersed in the sample centrifuge tube. The blackboard signal is acquired first, i.e., the light source is turned off or blocked, to measure the dark current / background noise in the current environment; whiteboard calibration is performed using air as a standard, and the reference spectrum of the whiteboard should be acquired before each sample reflectance measurement; then the sample to be tested (i.e., the simulated sample) is placed in the optical path, and an appropriate integration time is set for measurement. Furthermore, each sample is acquired three times, and the average value is taken as the final spectral data (i.e., the target spectral data, corresponding to…). Figure 3 (Near-infrared spectrum). The blackboard serves as the machine background to ensure stable and reliable spectral signals; in this application, the integration time, i.e., the time for collecting signal light from the light source after reflection (or transmission) through the sample, is set to 600 μs.

[0039] Therefore, the total ginsenoside content in American ginseng can be determined based on the target spectral data: Reference concentrations were determined using an Agilent 1260 Infinity ultra-high performance liquid chromatograph (UPLC) at room temperature. An Agilent XDB C18 column (4.6 mm × 150 mm, 5 μm) was used, with a mobile phase consisting of acetonitrile (A) and water (B). The gradient elution program was as follows: 0–20 min, 20% A; 20–35 min, 20%–35% A; 35–42 min, 35% A. The flow rate was set at 1.0 mL / min, the column temperature at 30 °C, the detection wavelength at 203 nm, and the injection volume at 10 μL. Each sample solution was measured independently three times, and the average value was used as the final reference concentration to establish the quantitative relationship between near-infrared spectral data and the actual saponin content.

[0040] Then, the target spectral data is preprocessed: first, white-black calibration is performed on the target spectral data to obtain the spectral transmittance (i.e., the relative transmittance of the sample at wavelength λ), which is then processed using the following formula: transmittance (1); Among them, the sample signal intensity is the light intensity received by the fiber optic probe after passing through the sample, at wavelength λ; the dark current signal intensity is the background noise signal output by the fiber optic probe under no-light conditions (such as turning off the light source or blocking the light), i.e., the blackboard signal collected; the reference signal intensity is the light intensity received by the fiber optic probe when passing through a blank reference (such as air or a transparent substrate), i.e., the whiteboard signal collected.

[0041] Next, transmittance is converted to absorbance. The absorbance conversion formula is: absorbance (2); Secondly, the Savitzky-Golay (SG) smoothing method is used to smooth the spectral data to be processed: the target spectral data after white-blackboard correction is regarded as a discrete signal sequence. Starting from the first data point, the sliding window is moved one data point at a time. Polynomial least squares fitting is performed in each sliding window, the fitting coefficient is calculated, and the fitted value of the center point is used to replace the original point value. After all data points are traversed, the smoothed spectral data sequence is obtained, which effectively removes high-frequency noise and retains important features such as peaks and troughs, so as to preserve the position and shape information of the original spectral peaks.

[0042] Furthermore, the spectral data to be processed is subjected to first-order derivative processing: the derivative of the smoothed absorbance data is obtained by using a difference algorithm (such as the central difference method or the SG derivative method), that is, the first-order derivative of the target spectral data after SG smoothing is obtained to further eliminate baseline drift and background interference, while enhancing the regions with prominent slope changes in the spectrum. This helps to extract small but significant differences in chemical composition, which is beneficial for component identification and differentiation.

[0043] Finally, the target spectral data after first-order derivative processing is subjected to min-max normalization: a linear transformation is performed on each set of spectral data (or each column) to scale the reflectance / absorbance values ​​of all wavelength points to the [0, 1] interval, resulting in time-sorted simulated spectral data for the current batch. This yields simulated spectral data collected at consecutive target times, ensuring consistent numerical scales between different samples and preventing feature value differences from affecting subsequent machine learning models. The same processing is then performed on other batches to obtain multiple batches of time-sorted simulated spectral data.

[0044] Notably, during data preprocessing, an independent processing strategy was adopted for each column chromatography elution batch to preserve its unique elution characteristics and concentration change trends, thereby enhancing the model's stability and generalization ability.

[0045] S300: The long short-term memory network combined model is trained in batches using the sliding time window mechanism on the training set in the simulated spectral data, so that the trained long short-term memory network combined model learns the mapping relationship between the time feature data in the simulated spectral data and the output target.

[0046] See also Figure 4 The combined long short-term memory network model (hereinafter referred to as the model) includes: two sequentially connected long short-term memory network sub-models (corresponding to...). Figure 4 The training dataset consists of an LSTM layer, a flattened layer, and a fully connected layer. The output targets include simulated and predicted concentrations of biomolecules, such as proteins, saponins, and antibodies. The sliding time window mechanism moves forward gradually over time, predicting data for the next time step based on the data in the current window, sliding only a small step before continuing prediction to traverse the entire training set. This mechanism is suitable for temporal supervised learning tasks, especially for elution processes where concentration changes are relatively gradual but exhibit clear trends. A time point in the time-sorted simulated spectral data constitutes a time step.

[0047] As an example, after obtaining the simulated spectral data, the third step is to train the pre-built model: In each batch, a sliding time window mechanism is used to divide the time series data in the simulated spectral data dataset (i.e., the first 6 batches mentioned above) into input subsequences (data in the time window) and target labels (data predicted during training). These subsequences are then input into the long short-term memory network combined model for training. This allows the model to learn the mapping relationship between the input data (time feature data in the simulated spectral data) and the output target (simulated predicted concentration of biomacromolecules). This enables the model to more accurately capture the dynamic trend of concentration changes over time during elution, improving the real-time performance and accuracy of the prediction.

[0048] It is understood that this application utilizes a sliding time window mechanism to train a combined long short-term memory network model using the training set in simulated spectral data. This enables the model to automatically learn the time series features in the spectral data, allowing it to learn the mapping relationship between the time features in the simulated spectral data and the output target. This model models the time-dependent features of the simulated spectral data during column chromatography elution, thereby enabling the model to accurately predict the concentration of ginsenosides at future times based on the mapping relationship during the prediction phase. This effectively and accurately predicts the behavior of the elution process.

[0049] In one implementation, each layer of the Long Short-Term Memory (LSTM) network submodel includes a target number of LSM networks connected sequentially. Each LSM network in the first layer of the LSM network submodel is connected to the LSM network at the same position in the second layer of the LSM network submodel. Each LSM network contains hidden units for extracting features from temporal feature data.

[0050] For demonstration purposes, please continue to refer to Figure 4 The Long Short-Term Memory (LSTM) network ensemble model contains two layers of LSTM sub-models to enhance the extraction capability of time series features. "2x128" indicates that the LSTM output has two directions, each with 128 (i.e., the target number) LSTM networks. In each LSTM sub-model, the 128 LSTM networks are sequentially connected, with each LSTM network in the first layer connected to the corresponding LSTM network in the second layer. Each LSTM network is a hidden layer unit.

[0051] Traditional static models struggle to effectively capture the patterns of concentration changes over time, impacting the accuracy and automation of concentration identification. However, LSTM (Long Short-Term Memory), an improved recurrent neural network (RNN), effectively extracts temporal dependencies from spectral sequences, enabling accurate prediction of concentration change trends, clear identification of elution start and end points, and significantly reducing human error. LSTM effectively controls the retention and forgetting of information over time by introducing gating mechanisms such as forget gates, input gates, and output gates, thus capturing long-term dependencies and making it suitable for processing dynamic changes and contextual information in time-series data.

[0052] like Figure 5 The diagram shows the structure of the LSTM hidden unit. The specific working mechanism of the LSTM hidden layer is as follows: At each time step, the current input data ( ) and the hidden state of the previous moment ( They are fed together into the LSTM cell. They are simultaneously fed into three different gate structures: the forget gate (…). ), Input gate ( ) and output gate ( These three gates will each output three "weight control signals" to adjust the processing method of the current information.

[0053] First, the forget gate determines how much of the memory information from the previous time step should be retained; it outputs a control signal to adjust the memory value from the previous time step. Next, the input gate determines which information from the current input and the previous hidden state is "new important content" and can be added to the current memory. Simultaneously, a module generates "candidate information," which is new content that may be written into the memory. The output of the input gate controls how much of this candidate information is written. The results of these two parts (the forgotten old memory and the selected new memory to be written) are merged to form the updated state at the current time step. Then, the output gate stage is entered, which determines which parts should be output as the current hidden state based on the updated memory. Finally, the output hidden state not only serves as the output result for this time step but is also passed to the next time step as part of the loop input, realizing information transfer over time.

[0054] In one implementation method, please refer to Figure 6 Step S300: Using a sliding time window mechanism, the training set from the simulated spectral data is used to train the Long Short-Term Memory Network Combination Model in batches, so that the trained Long Short-Term Memory Network Combination Model learns the mapping relationship between the simulated spectral data and the output target, including: S310. In each batch, the long short-term memory network combination model is used to predict the training set corresponding to the current time window, so as to obtain the simulated predicted concentration of biomolecules at the next moment after the current time window. S320. Predict the training set corresponding to the next time window selected after moving a preset distance, and obtain the simulated predicted concentration of biomolecules at the next time window, until the training set at all times is predicted, and obtain the simulated predicted concentration sequence. S330. Train the simulated predicted concentration sequence until the loss function of the long short-term memory network combined model reaches the target threshold, so that the trained long short-term memory network combined model can learn the mapping relationship between the time feature data in the simulated spectral data and the simulated predicted concentrations of the corresponding biomolecules.

[0055] Each time window has the same preset time step length, such as 3 time steps; the time feature data consists of spectral data sorted by time.

[0056] As an example, the specific implementation process of the sliding time window mechanism is as follows: The system sets the time window length to 3 time steps (i.e., the preset time step length), and slides forward 1 time step (i.e., the preset distance) each time, continuously extracting 3 adjacent sets of feature values ​​from the training set of the simulated spectral data as an input subsequence. This is used to simulate and predict the concentration of ginsenosides at the fourth time point (i.e., the moment after the current time window). This creates a supervised learning sample set where "the first three steps predict the next step." The time window is continuously moved forward, traversing the entire dataset to form sufficient training samples for modeling and training the deep learning model.

[0057] Since the concentration of ginsenosides in the first 12 of the 25 samples collected in each batch during the sample collection phase is 0, the model training phase mainly focuses on the 13 samples that contain concentration (belonging to the effective elution phase). This application uses the last 16 samples out of the 25 samples for model training. Following the sliding time window mechanism in this application (the time window length is 3 time steps, sliding forward 1 time step each time), 13 sub-sequences can be extracted from the last 16 samples in each batch. These sub-sequences are used for model training; that is, 13 sub-sequences are generated for each batch through the sliding time window mechanism. Therefore, the sub-sequences generated in the first 6 batches of the 8 simulated sample collections, i.e., the simulated predicted concentration sequences, are all used for model training.

[0058] Then, the Long Short-Term Memory (LSTM) network combined model is trained on the simulated predicted concentration sequence. At this point, the input to the LSM network combined model is a three-dimensional tensor, namely the simulated predicted concentration sequence (corresponding to...). Figure 4 The shape of the near-infrared spectrum after constructing the subsequence is [3, 3, 147], corresponding to the batch size (batch_size=3), time window length (seq_len=3), and input feature dimension (input_size=147), respectively. The 3D tensor passes through an LSTM layer, then a flattening layer, outputting a 384-dimensional vector. This is followed by a fully connected layer, outputting a 64-dimensional vector. The model's final output is the hidden state of the last time step in the sequence. This hidden state is then feature-mapped by the next fully connected neural network layer (i.e., the fully connected layer), outputting a predicted value for the target variable—total saponin concentration—achieving regression prediction of the concentration in the current time period. The batch size is set to 3 to accommodate training requirements in small sample scenarios.

[0059] Furthermore, during model training, the Adam (Adaptive Moment Estimation) optimizer or SGD (Stochastic Gradient Descent) optimizer is used in a GPU environment to accelerate training and continuously adjust model parameters to speed up the reduction of the loss function value, i.e., speed up convergence and improve training stability, until the final model's loss function calculation result reaches the target threshold, such as 0.001, so that the trained long short-term memory network combined model can learn the mapping relationship between the temporal feature data in the simulated spectral data and the simulated predicted concentrations of biomacromolecules.

[0060] S400: Using the trained long short-term memory network combined model, based on the mapping relationship and the actual spectral data collected during the column chromatography elution process at the target time, the actual predicted concentration of biomacromolecules at the time following the target time is predicted, so as to determine the key nodes in the column chromatography elution process.

[0061] The key milestones include the elution start point and the elution end point. The elution start point is the first detection of a ginsenoside signal after the start of eluent collection, or the first time the ginsenoside concentration in the eluent reaches a preset threshold. The elution end point is when the detection signal returns to the baseline level and the ginsenoside concentration decreases below the set threshold within multiple consecutive collection cycles.

[0062] Demonstratively, after training a combined long short-term memory network model, the trained combined long short-term memory network model is used to predict actual elution behavior during actual column chromatography elution: First, the near-infrared fiber optic probe of the near-infrared spectrometer is inserted into the path of the industrial column chromatography eluent to collect data in real time. During the actual column chromatography elution process, the two time steps before the current time are taken together with the current time as the target time. The actual spectral data collected at the target time (e.g., three time steps) are input into the trained long short-term memory network combined model to predict the actual predicted concentration of biomolecules at the time step after the target time, so as to predict the elution start and elution end points.

[0063] It is understood that this application employs near-infrared spectroscopy technology for non-contact real-time monitoring of column chromatography eluent. By acquiring spectral data online through a fiber optic probe, dynamic and non-destructive detection of the elution process is achieved, preserving the active structure of the biological sample within the chromatography column. The active structure is then input into a trained long short-term memory network (LSTM) model for detection, enabling timely and intelligent identification of key nodes in the elution process. This allows operators to more intuitively and accurately grasp the elution timing of target components, effectively improving the stability and production efficiency of the elution process.

[0064] In one implementation method, the method for predicting the elution behavior further includes: The validation set from the simulated spectral data is input into the trained long short-term memory network ensemble model to calculate the evaluation index, and the structure or hyperparameters of the trained long short-term memory network ensemble model are adjusted based on the calculation results.

[0065] The evaluation indicators include: coefficient of determination, mean absolute error, root mean square error, and relative prediction bias.

[0066] As an example, after obtaining the trained combined long short-term memory network model, evaluation metrics can be used for model validation or evaluation to assess the model's performance in the task of predicting ginsenoside content during column chromatography: The validation set from the simulated spectral data, specifically the 7th batch out of the first 8 batches, is input into the trained Long Short-Term Memory (LSTM) network ensemble model. Prediction is performed using a sliding time window mechanism, predicting the concentration for the next time step based on the validation data every three time steps. The predicted values ​​at multiple time steps are compared with the actual values ​​at the corresponding time steps, and the calculations are performed using the respective formulas for the evaluation metrics (including coefficient of determination, mean absolute error, root mean square error, and relative prediction bias). The structure or hyperparameters of the trained LSM network ensemble model are then adjusted based on the calculation results, and the training and validation process is repeated. A high R² indicates a good model fit; low MAE and RMSE indicate a small overall prediction error, demonstrating good stability and practicality in real-world applications; RPD > 4 indicates excellent predictive ability. In summary, the model demonstrates good predictive ability and stability, indicating its applicability in real-world production environments.

[0067] R² (coefficient of determination) measures the goodness of fit of a model to the data; a value closer to 1 indicates a better fit. The formula for calculating the coefficient of determination is: (3); MAE (Mean Absolute Error) measures the average absolute difference between predicted and actual values, providing a direct reflection of the model's average prediction bias in practical applications. The formula for calculating MAE is: (4); RMSE (Root Mean Square Error) reflects the average squared value of prediction errors. It is more sensitive to larger deviations and is suitable for evaluating the overall prediction accuracy of a model. The formula for calculating RMSE is: (5); RPD (Relative Prediction Deviation) is the ratio of the standard deviation of the sample reference values ​​to the prediction error (RMSE). It measures a model's predictive ability relative to the range of data fluctuations. A higher RPD value indicates better model stability and applicability. The formula for calculating relative prediction deviation is: (6); In formulas (3)-(6), Indicates the first The true value of each sample Indicates the first The predicted value for each sample, The mean of the true values. is the total number of test samples, where the test samples are the validation set.

[0068] It is understandable that evaluating the predictive accuracy and stability of a model through a validation set can improve its reliability and accuracy in practical applications.

[0069] In one implementation method, please refer to Figure 7 The model utilizes a sliding time window mechanism to train the Long Short-Term Memory network ensemble model in batches using the training set from simulated spectral data. This process also includes: S301. After each prediction is completed on the training set at all times in the current batch, the determination coefficient is calculated using the validation set. S302. If the coefficient of determination remains unchanged throughout the training process of consecutive target batches, the training process ends; otherwise, the training process continues until the training set of all batches at all times is predicted.

[0070] The training process for the current batch ends when all training sets at all times are predicted.

[0071] As an example, an early stopping mechanism is introduced during the training of a combined long short-term memory network model: Each time a training cycle ends after the Long Short-Term Memory Network combined model completes predictions for all time points of the current batch on the training set, the coefficient of determination is calculated using the validation set.

[0072] However, if the coefficient of determination remains unchanged throughout the training of consecutive target batches—for example, if the coefficient of determination fails to improve or change further over four consecutive training epochs—the training process should be terminated early. This is because it indicates that the model has reached its best performance on the current dataset, and continuing training might lead to overfitting. Therefore, it is not necessary to wait until all training sets have been completed before ending the training process.

[0073] However, if the coefficient of determination keeps changing or increasing during training, the training process is only completed after the training set has completed predictions in all batches and at all times.

[0074] It is understood that by introducing an early stopping mechanism in this application, overfitting can be prevented, computational resources can be saved, the training process can be dynamically adjusted, and the generalization ability and experimental efficiency of the model can be improved.

[0075] Please see Figure 8 This application provides a system for predicting elution behavior, comprising: The acquisition module 10 is used to acquire the target liquid precipitated during the simulated column chromatography elution process multiple times at equal time intervals, and obtain multiple batches of simulated acquisition samples sorted by time.

[0076] The preprocessing module 20 is used to perform near-infrared spectral acquisition and preprocessing on all simulated samples in each batch to obtain multiple batches of simulated spectral data sorted by time.

[0077] Training module 30 is used to train the Long Short-Term Memory Network Combination Model on the training set in the simulated spectral data in batches using a sliding time window mechanism, so that the trained Long Short-Term Memory Network Combination Model can learn the mapping relationship between the temporal feature data in the simulated spectral data and the output target.

[0078] The prediction module 40 is used to use the trained long short-term memory network combined model to predict the actual predicted concentration of biomacromolecules at the time following the target time in the column chromatography elution process, based on the mapping relationship and the actual spectral data collected at the target time. This is to determine the key nodes in the column chromatography elution process.

[0079] By way of example, the system of this embodiment corresponds to the method for predicting elution behavior in the above embodiments. The options in the above embodiments are also applicable to this embodiment, so they will not be described again here.

[0080] Further, please refer to Figure 9 The embodiment also provides a real-time detection system for the column chromatography elution process, including: online acquisition units connected in sequence (corresponding to...). Figure 9 Part a), near-infrared spectroscopy acquisition module (corresponding to Figure 9 Part b of the document), the back-end processing system (corresponding to...) Figure 9 Part C in the middle) and the user visualization module (corresponding to Figure 9 (part d in the text).

[0081] The online acquisition unit is used to collect the target liquid precipitated during the simulated column chromatography elution process at equal time intervals to obtain simulated samples. This part simulates an industrial production scenario, with a flow cell and a near-infrared fiber optic probe installed at the eluent outlet. This allows the eluent to fully contact the probe during discharge, thereby achieving continuous, non-destructive, and real-time spectral acquisition of the sample. This ensures stable and effective acquisition of near-infrared signals, which is beneficial to the accuracy of subsequent data analysis.

[0082] Near-infrared spectroscopy acquisition module: By connecting to a near-infrared spectrometer, it acquires raw near-infrared spectral information in the wavelength range of 900~1700 nm (target wavelength range), obtains target spectral data, and transmits it to the background processing system in real time to obtain near-infrared characteristic curves reflecting changes in eluent composition, which serve as the basis for subsequent judgment and can be displayed by the user's visualization module.

[0083] The backend processing system includes a data preprocessing module and a deployed LSTM deep learning model. The data preprocessing module performs preprocessing operations on the target spectral data to improve data quality. Based on the preprocessed target spectral data at the target time (three time steps), it predicts the actual concentration of biomolecules at the next time step after the target time. In other words, the preprocessed target spectral data is input into the trained LSTM model to predict the total ginsenoside concentration at the next future time point.

[0084] The alarm system issues a warning signal after identifying key points. The system determines the start time of elution by monitoring the slope of the concentration rise curve in the near-infrared characteristic curve; when the concentration curve gradually decreases and tends to stabilize, with the slope approaching zero, it is determined that elution has ended. Both key points can trigger an alarm prompt, assisting the operator in accurately controlling the elution process.

[0085] User visualization module: The user terminal has a visualization interface that displays the concentration change curve during the elution process in real time, and issues early warning signals by identifying key nodes through the system.

[0086] The real-time detection system for the column chromatography elution process has good scalability and adaptability, and can be widely applied to the column chromatography separation process of various Chinese medicinal materials, providing technical support for process analysis and quality control in the modern production of Chinese medicine.

[0087] Then please see Figure 10 The graph shows a comparison of the target component concentration over time during column chromatography elution. The horizontal axis represents elution time points (0-12), corresponding to the eluent elution process; the vertical axis represents the target component concentration, specifically the total ginsenoside concentration (mg / g) in this application. The solid black line represents the actual concentration value, and the dashed gray line represents the predicted concentration value. Both show an initial increase followed by a decrease, reaching a peak at time point "2". The initial increase is due to sufficient contact between the eluent and the stationary phase, causing the target component to continuously desorb from the stationary phase and enter the mobile phase, resulting in a gradual increase in concentration. The subsequent decrease is because most of the target component on the stationary phase has been eluted, reducing the residual amount and enhancing the diffusion effect, leading to a continuous decrease in the concentration in the eluent. Figure 10 It can be seen that the trend of the model's predicted values ​​is basically consistent with the actual values, especially in the concentration change range in the middle and later stages, which shows that the model has good fitting ability and time series trend capture ability.

[0088] This application also provides a terminal device, exemplary of which includes a processor and a memory, wherein the memory stores a computer program, and the processor, by running the computer program, causes the terminal device to perform the functions of the various modules in the above-described method for predicting elution behavior or the above-described system for predicting elution behavior.

[0089] The processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, including at least one of a Central Processing Unit (CPU), Graphics Processing Unit (GPU), Network Processor (NP), Digital Signal Processor (DSP), Application-Specific Integrated Circuit (ASIC), Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application.

[0090] The memory can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The memory is used to store computer programs, and the processor can execute the computer programs accordingly after receiving execution instructions.

[0091] This application also provides a computer-readable storage medium for storing the computer program used in the aforementioned terminal device. For example, the computer-readable storage medium may include, but is not limited to, various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0092] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative; for example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that, as an alternative implementation, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0093] In addition, the functional modules or units in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0094] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, 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 smartphone, personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0095] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for predicting elution behavior, characterized in that, include: The target liquid precipitated during the simulated column chromatography elution process was collected multiple times at equal time intervals to obtain multiple batches of simulated samples sorted by time. Near-infrared spectra were collected from all simulated samples in each batch and preprocessed to obtain multiple batches of simulated spectral data sorted by time. The training set of the simulated spectral data is used in batches to train the Long Short-Term Memory Network Combination Model, so that the trained Long Short-Term Memory Network Combination Model learns the mapping relationship between the temporal feature data and the output target in the simulated spectral data. Using the trained long short-term memory network combined model, based on the mapping relationship and the actual spectral data collected during the column chromatography elution process at the target time, the actual predicted concentration of biomacromolecules at the time following the target time during the column chromatography elution process is predicted, so as to determine the key nodes in the column chromatography elution process.

2. The method for predicting elution behavior according to claim 1, characterized in that, The output targets include: simulated and predicted concentrations of biomacromolecules; The step of using a sliding time window mechanism to train the Long Short-Term Memory (LSTM) network ensemble model in batches using the training set from the simulated spectral data, so that the trained LSM network ensemble model learns the mapping relationship between the simulated spectral data and the output target, includes: In each batch, the long short-term memory network combination model is used to predict the training set corresponding to the current time window to obtain the simulated predicted concentration of the biomolecule at the next moment after the current time window. The training set corresponding to the next time window selected after moving a preset distance is predicted to obtain the simulated predicted concentration of the biomolecule at the next time window, until the training set at all times is predicted to obtain the simulated predicted concentration sequence. The simulated predicted concentration sequence is trained until the loss function of the long short-term memory network combined model reaches the target threshold, so that the trained long short-term memory network combined model learns the mapping relationship between the time feature data in the simulated spectral data and the simulated predicted concentration of the biomacromolecule. Each time window has the same preset time step length; the time feature data consists of spectral data sorted by time; the biomacromolecules include proteins, saponins, and antibodies; the key nodes include elution start point and elution end point.

3. The method for predicting elution behavior according to claim 1, characterized in that, The combined long short-term memory network model includes: a two-layer long short-term memory network sub-model, a flattened layer, and a fully connected layer connected in sequence; Each layer of the long short-term memory network sub-model includes a target number of long short-term memory networks connected sequentially. Each long short-term memory network in the first layer of the long short-term memory network sub-model is connected to the long short-term memory network at the same position in the second layer of the long short-term memory network sub-model. Each Long Short-Term Memory (LSTM) network contains hidden units for extracting features from the temporal feature data.

4. The method for predicting elution behavior according to claim 1, characterized in that, The preprocessing includes: white-blackboard correction, absorbance conversion, SG smoothing, first derivative processing, and maximum-minimum normalization. The process involves acquiring near-infrared spectra of all simulated samples in each batch and preprocessing them to obtain multiple batches of simulated spectral data sorted by time, including: The target spectral data are obtained by collecting spectral data of all the simulated samples in the target wavelength range using a near-infrared spectrometer. The target spectral data is sequentially subjected to white-black correction, absorbance conversion, SG smoothing, first derivative processing, and maximum-minimum normalization processing to obtain the multiple batches of simulated spectral data sorted by time. The time-sorted simulated spectral data refers to simulated spectral data collected at consecutive target times.

5. The method for predicting elution behavior according to claim 2, characterized in that, Also includes: The validation set from the simulated spectral data is input into the trained long short-term memory network ensemble model to calculate the evaluation index, and the structure or hyperparameters of the trained long short-term memory network ensemble model are adjusted according to the calculation results. The evaluation indicators include: coefficient of determination, mean absolute error, root mean square error, and relative prediction bias.

6. The method for predicting elution behavior according to claim 5, characterized in that, The step of using a sliding time window mechanism to train the Long Short-Term Memory network ensemble model in batches using the training set from the simulated spectral data also includes: After each prediction is completed on the training set at all times in the current batch, the determination coefficient is calculated using the validation set. If the determination coefficient remains unchanged throughout the training process of consecutive target batches, the training process ends; otherwise, the training process continues until the training set completes prediction for all batches at all times. The completion of prediction of the training set at all times in the current batch is considered the end of the training process for the current batch.

7. A system for predicting elution behavior, characterized in that, include: The acquisition module is used to acquire the target liquid precipitated during the simulated column chromatography elution process multiple times at equal time intervals, and obtain multiple batches of simulated acquisition samples sorted by time. The preprocessing module is used to perform near-infrared spectral acquisition and preprocessing on all simulated samples in each batch to obtain multiple batches of simulated spectral data sorted by time. The training module is used to train the Long Short-Term Memory Network Combination Model on the training set in the simulated spectral data in batches using a sliding time window mechanism, so that the trained Long Short-Term Memory Network Combination Model learns the mapping relationship between the temporal feature data in the simulated spectral data and the output target. The prediction module is used to use the trained long short-term memory network combined model to predict the actual predicted concentration of biomacromolecules at the time following the target time in the column chromatography elution process, based on the mapping relationship and the actual spectral data collected at the target time, so as to determine the key nodes in the column chromatography elution process.

8. A real-time detection system for column chromatography elution process, characterized in that, include: The system consists of an online data acquisition unit, a near-infrared spectroscopy acquisition module, a back-end processing system, and a user visualization module, connected sequentially. The online acquisition unit is used to collect the target liquid precipitated during the simulated column chromatography elution process at equal time intervals to obtain simulated samples. The near-infrared spectroscopy acquisition module is used to collect spectral data of the target wavelength range in the simulated sample, obtain target spectral data, and generate a near-infrared characteristic curve based on the target spectral data so that the user visualization module can display it. The background processing system is used to preprocess the target spectral data and predict the actual concentration of biomacromolecules at the next time step after the target time based on the preprocessed target spectral data at the target time. It is also used to issue an early warning signal when the key nodes in the column chromatography elution process are determined based on the near-infrared characteristic curve.

9. A terminal device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method for predicting elution behavior as described in any one of claims 1-6.

10. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the steps of the method for predicting elution behavior as described in any one of claims 1-6.

Images