Real-time modulation multiplexing tandem mass spectrometry detection method

Abstract

The invention discloses a real-time modulation multiplexing tandem mass spectrometry detection method which comprises the following steps: loading a first sample to be detected to a first sample introduction channel, and loading a second sample to be detected to a second sample introduction channel; starting a first liquid chromatography separation column to generate a first separation flow; dynamically predicting a target fraction time window set based on the real-time chromatographic data of the first separation flow; starting a second liquid chromatographic separation column at the starting moment of the first non-target fraction section to generate a second separation flow when it is judged that the second liquid chromatographic separation column enters the non-target fraction section; performing dynamic prediction on the second separation flow to generate a target fraction time window set of the second separation flow; if the current moment t belongs to any time period Wi in the target fraction time window set of the first separation flow, introducing the first separation flow into a mass spectrometer; and if the current moment t belongs to any time period Vj in the target fraction time window set of the second separation flow, introducing the second separation flow into the mass spectrometer. According to the invention, the mass spectrum detection efficiency and reliability can be obviously improved.

Description

Technical Field

[0001] This invention relates to the field of mass spectrometry detection technology. More specifically, this invention relates to a real-time modulated multiplexed tandem mass spectrometry detection method. Background Technology

[0002] In the field of mass spectrometry, traditional liquid chromatography-mass spectrometry (LC-MS) technology has long been limited by the dual bottlenecks of detection throughput and equipment utilization. Existing single-channel LC-MS systems are not only cumbersome and prone to significant operational errors, but also require frequent column changes for different detection items, resulting in the mass spectrometer being idle in non-target fractions and insufficient sample throughput per unit time. While CN118112138A achieves time-dimension reuse through dual-injection separation modules and valve module switching, existing technology still has significant shortcomings: the trigger control module uses a preset time to start the separation module, without dynamic adjustment based on real-time chromatographic data. When sample retention times fluctuate (e.g., due to matrix differences between different batches of samples), it easily leads to misalignment between the target fraction and the mass spectrometry detection window; when multiple channels operate in parallel, there is a lack of prediction of chromatographic peak elution times, relying solely on fixed-sequence flow path switching, resulting in detection conflicts during channel overlap periods and compromised data integrity.

[0003] Therefore, it is necessary to design a technical solution that can overcome the above-mentioned defects. Summary of the Invention

[0004] One objective of this invention is to provide a real-time modulated multiplexed tandem mass spectrometry detection method that can significantly improve the efficiency and reliability of mass spectrometry detection.

[0005] To achieve these objectives and other advantages of the present invention, according to one aspect of the present invention, the present invention provides: S1: loading a pretreated first sample to be tested into a first injection channel, and loading a pretreated second sample to be tested into a second injection channel, wherein the first injection channel and the second injection channel are physically isolated; S2: starting a first liquid chromatography column to separate the first sample to be tested, generating a first separation stream; S3: based on the real-time chromatographic data of the first separation stream, dynamically predicting its target fraction time window set through a machine learning model, wherein the target fraction time window set includes n consecutive time periods W1 to W... n W for each time period i From start time t i -start and end times t i-end definition; S4: Real-time monitoring of the chromatographic signal change rate of the first separated stream. When the absolute value of the signal change rate does not exceed 0.05 AU / s for 5 consecutive seconds, it is determined that the stream has entered a non-target distillation segment. At the beginning of the first non-target distillation segment, the second liquid chromatography separation column is started to process the second sample to be tested, generating the second separated stream; S5: Perform dynamic prediction of S3 on the second separated stream to generate its target distillation time window set, which contains m consecutive time periods V1 to V m V for each time period j From start time t j -start and end times t j -end definition; S6: If the current time t belongs to any time period W in the target fraction time window set of the first separation stream. i The first separated stream is fed into the mass spectrometer; if the current time t belongs to any time period V in the target fraction time window set of the second separated stream... j The second separation stream is fed into the mass spectrometer; if the current time t does not belong to the time window of any target fraction, the output stream is switched to the waste liquid channel.

[0006] Further, S3 includes: S3.1: Obtaining 200 to 300 sets of chromatographic-mass spectrometry (GC-MS) data of different chemical properties of test samples from the historical detection database. Each set of data includes retention time axis, UV absorption signal intensity, and mass spectrometry characteristic precursor ion intensity; S3.2: Constructing a transfer learning model framework, which includes a basic feature extraction layer, a distribution adaptation module, and an output layer. The basic feature extraction layer uses a convolutional neural network to process the chromatographic signal, with the input dimension being the timestamp multiplied by the signal intensity, and the output being a 128-dimensional feature vector. The distribution adaptation module aligns the feature distribution of different analytes using the maximum mean difference algorithm, with the maximum mean difference loss function weight set to 0.35-0.45. The output layer is a fully connected neural network structure; S3.3: For the first test sample, acquiring a 20-second real-time chromatographic signal segment every 2 seconds within 0 to 60 seconds after separation starts; inputting the real-time chromatographic signal segment into the transfer learning model in S3.2, and outputting an initial prediction window set P1 to P2. k S3.4: When the retention time of the first separated stream reaches the start time t of the first window P1 in the initial prediction window set. p1 At -start, perform mass spectrometry characteristic precursor ion intensity verification: at t p1 -start 3 seconds before t p1 If the intensity of the characteristic precursor ion of the target compound exceeds 1×10⁻⁶ within 3 seconds after starting, 4 If the counts are positive, then P1 is confirmed to be valid; if the characteristic precursor ion strength does not exceed 5 × 10⁻⁶, then P1 is valid. 3 If counts are found, then P1 is marked as invalid and incremental training of the model is activated.

[0007] Furthermore, the incremental training of the model includes: adding real-time chromatographic signal fragments corresponding to marked invalid windows to the training dataset; freezing the parameters of the basic feature extraction layer; updating the output layer weights, with 15-20 training iterations; and using the updated transfer learning model to re-predict subsequent windows to generate the final target distillation time window set W1 to W2. n .

[0008] Furthermore, when the first window P1 in the initial prediction window set is marked as invalid, the following operations are performed: reference chromatographic data with the same target compound category as the first test sample is retrieved from the historical database, and all target fraction time window parameters in the reference data are extracted as the default window set; before the transfer learning model completes incremental training, the default window set is overlaid onto all windows to be verified in the initial prediction window set except for the first window; after the incremental training is completed, the default window set is immediately replaced with the newly generated final target fraction time window set.

[0009] Furthermore, the determination of the same target compound category is achieved through the following steps: obtaining the precise molecular weight data, measured lipid-water partition coefficient data, and topological polar surface area data of the target compound in the first test sample; inputting the aforementioned molecular weight data, lipid-water partition coefficient data, and topological polar surface area data into a pre-trained compound classification model to generate a sample feature vector; calculating the similarity value between the sample feature vector and the feature vectors of all reference compounds in the historical database; when there is a reference compound with a similarity value greater than or equal to 0.9, determining that the reference compound belongs to the same target compound category as the first test sample, and selecting the reference compound data source with the highest similarity value.

[0010] Further, S4 includes: S4.1: Real-time acquisition of UV absorption signal intensity data of the first separated stream; S4.2: Calculation of the absolute value of the continuous signal change rate, wherein the signal change rate is defined as the absolute value of the difference between the absorbance values ​​of adjacent sampling points divided by a 0.1-second time interval; S4.3: When the absolute value of the signal change rate is detected to be no more than 0.05 AU / s for 50 consecutive data points, mass spectrometry verification is activated; S4.4: The mass spectrometry verification performs the following operations: Acquiring mass spectrometry characteristic precursor ion intensity data within the interval from 2 seconds before to 2 seconds after the trigger time; If the characteristic precursor ion intensity is less than 1% of the maximum intensity of the current chromatographic peak, it is confirmed that the non-target distillation segment has been entered; S4.5: At the start time of the first non-target distillation segment that meets the conditions of S4.4, a start command is sent to the second liquid chromatography separation column.

[0011] Furthermore, determining the maximum intensity of the current chromatographic peak in S4.4 includes: using the moment of activation of mass spectrometry verification as the trigger time; and setting a dynamic backtracking time interval [t]. start , ttrigger , where the starting time t start is calculated according to the following formula: t start = t trigger - min(120, 3×W avg ); where W avg is the average chromatographic peak width of the target compound in the current liquid chromatography system, and its value is obtained through the calibration experiment in the system initialization stage. The calibration method is as follows: perform chromatographic separation on at least 5 standard target compounds, and calculate the average of the half-peak widths of the main chromatographic peaks of each compound; extract the maximum value of the ultraviolet absorption signal intensity within the time interval [tstart, ttrigger], denoted as I max , and take I max as the maximum intensity output of the current chromatographic peak.

[0012] Furthermore, S6 also includes: when the current time t belongs to any window W i in the target fraction time window set of the first separation flow and any window V j in the target fraction time window set of the second separation flow at the same time, perform the following operations: compare the start time stamp T1 of the first liquid chromatography separation column with the start time stamp T2 of the second liquid chromatography separation column; if T1 < T2, introduce the first separation flow into the mass spectrometer, and at the same time switch the second separation flow to the waste liquid channel; if T2 < T1, introduce the second separation flow into the mass spectrometer, and at the same time switch the first separation flow to the waste liquid channel.

[0013] The present invention at least includes the following beneficial effects: Through the dual injection channels and the real-time modulated multiplexing mechanism, the present invention significantly improves the mass spectrometry detection efficiency and reliability. Based on the non-target section determination technology of the real-time chromatographic signal change rate, it accurately captures the idle time period in liquid chromatography separation, starts the second separation column at the starting moment of the first non-target section, realizes the time-dimensional multiplexing of the dual separation flows, and greatly improves the use efficiency of the mass spectrometer. The mechanism of the machine learning model for dynamically predicting the target fraction time window breaks through the limitations of the traditional fixed window, adaptively generates the target fraction interval of continuous time periods according to the real-time chromatographic data, and makes the capture of the target compound by the mass spectrometer more targeted. The flow path switching logic accurately introduces the target separation flow into the mass spectrometer through the time window matching mechanism, switches to the waste liquid channel during the non-target period to reduce background interference; when there is a time window conflict between the two channels, the priority is determined according to the start sequence of the separation columns to ensure the orderliness and data integrity of the detection process, and provides an efficient and reliable solution for the tandem mass spectrometry detection of high-throughput and multi-type samples.

[0014] Other advantages, objectives and features of the present invention will be partially reflected by the following description, and partially will also be understood by those skilled in the art through the research and practice of the present invention. Attached Figure Description

[0015] Figure 1 This is a flowchart of one embodiment of this application. Detailed Implementation

[0016] The present invention will now be described in further detail so that those skilled in the art can implement it based on the description.

[0017] It should be understood that terms such as "having," "comprising," and "including" used in the embodiments of this application do not exclude the presence or addition of one or more other elements or combinations thereof. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationship and movement of components in a specific posture. If the specific posture changes, the directional indication will also change accordingly. When an element is referred to as "fixed to" or "set on" another element, it can be directly on the other element or may have an intervening element present. When an element is referred to as "connected to" another element, it can be directly connected to the other element or indirectly connected to the other element through an intervening element. Descriptions involving "first," "second," etc., in the embodiments of this application are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features.

[0018] It should be noted that the technical solutions of the various embodiments of this application can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by this application.

[0019] like Figure 1 As shown, embodiments of this application provide a real-time modulated multiplexed tandem mass spectrometry detection method comprising: loading a pretreated first analyte sample into a first injection channel, loading a pretreated second analyte sample into a second injection channel, wherein the first and second injection channels are physically isolated; starting a first liquid chromatography column to separate the first analyte sample, generating a first separation stream; and dynamically predicting a target fraction time window set based on real-time chromatographic data of the first separation stream using a machine learning model, wherein the target fraction time window set includes n consecutive time periods W1 to W2. n W for each time period i From start time t i -start and end times t i-end definition; Real-time monitoring of the chromatographic signal change rate of the first separated stream. When the absolute value of the signal change rate does not exceed 0.05 AU / s for 5 consecutive seconds, it is determined that the stream has entered a non-target distillation segment. At the start of the first non-target distillation segment, the second liquid chromatography separation column is started to process the second sample to be tested, generating a second separated stream. Dynamic prediction is performed on the second separated stream to generate its target distillation time window set, which contains m consecutive time periods V1 to V... m V for each time period j From start time t j -start and end times t j -end definition; if the current time t belongs to any time period W in the target fraction time window set of the first separation stream. i The first separated stream is fed into the mass spectrometer; if the current time t belongs to any time period V in the target fraction time window set of the second separated stream... j The second separation stream is fed into the mass spectrometer; if the current time t does not belong to the time window of any target fraction, the output stream is switched to the waste liquid channel.

[0020] For example, the pretreated sample refers to a biological sample that has undergone solid-phase extraction or protein precipitation. The first and second injection channels can be independent flow paths in an Agilent 1290 Infinity II liquid chromatograph, using stainless steel tubing with an inner diameter of 0.1 mm, physically isolated by a two-position six-way valve. The first liquid chromatography column can be a Waters ACQUITY UPLC BEH C18 column (2.1 × 100 mm, 1.7 μm). During separation, mobile phase A is 0.1% formic acid aqueous solution, and mobile phase B is acetonitrile, eluted by a binary pump at a flow rate of 0.3 mL / min gradient. The machine learning model runs on an industrial computer equipped with an NVIDIA Tesla T4 graphics card, acquiring chromatographic data in real time at a frequency of 10 Hz. The predicted time window for the target fraction can be a continuous interval of 8 seconds, 12 seconds, or 16 seconds. When monitoring the rate of change of the signal, the sampling interval is 0.1 seconds. The absolute difference between adjacent absorbance values ​​(e.g., at a wavelength of 214 nm) is divided by the time interval. When 50 consecutive data points (corresponding to 5 seconds) are all ≤0.05 AU / s, the second separation column is triggered. The second separation column can be the same model as the first column. The mass spectrometer can be a BrukermaXis 4G QTOF. Flow path switching is performed via a VICI Valco electric six-way valve, and the waste liquid channel is connected to a PTFE collection bottle.

[0021] In existing technologies, traditional mass spectrometry detection typically employs single-channel sequential injection. When chromatographic separation occurs in a non-target fraction segment, the mass spectrometer remains idle due to the lack of sample inflow, resulting in a limited sample detection capacity per unit time. This embodiment, however, utilizes a dual-injection channel with physical isolation design. By initiating the separation process of the other channel during the non-target fraction segment's time period, and combining this with machine learning-based dynamic prediction of the time window to precisely control the flow path switching, the mass spectrometer can process more samples per unit time, effectively increasing the throughput of the detection system.

[0022] In another embodiment, S3 includes: acquiring 200 to 300 sets of chromatographic-mass spectrometry (GC-MS) data of test samples with different chemical properties from a historical detection database, each set of data including retention time axis, UV absorption signal intensity, and mass spectrometry characteristic precursor ion intensity; constructing a transfer learning model framework, which includes a basic feature extraction layer, a distribution adaptation module, and an output layer; the basic feature extraction layer uses a convolutional neural network to process the chromatographic signal, with the input dimension being the timestamp multiplied by the signal intensity, and the output being a 128-dimensional feature vector; the distribution adaptation module aligns the feature distribution of different test analytes using a maximum mean difference algorithm, with the maximum mean difference loss function weight set to 0.35-0.45; the output layer is a fully connected neural network structure; for the first test sample, acquiring a real-time chromatographic signal fragment with a duration of 20 seconds every 2 seconds within 0 to 60 seconds after separation starts; inputting the real-time chromatographic signal fragment into the transfer learning model, and outputting an initial prediction window set P1 to Pk; when the retention time of the first separated stream reaches the start time t of the first window P1 in the initial prediction window set... p1 At -start, perform mass spectrometry characteristic precursor ion intensity verification: at t p1 -start 3 seconds before t p1 If the intensity of the characteristic precursor ion of the target compound exceeds 1×10⁻⁶ within 3 seconds after starting, 4 If the counts are positive, then P1 is confirmed to be valid; if the characteristic precursor ion strength does not exceed 5 × 10⁻⁶, then P1 is valid. 3 If counts are found, then P1 is marked as invalid and incremental training of the model is activated.

[0023] For example, the historical detection database is stored in an Oracle database, containing samples of different categories such as antibiotics and hormones. The retention time axis for each data set is at 0.1-second intervals. The UV signal intensity is acquired at a wavelength of 254 nm, and the mass spectrometry precursor ion intensity is acquired using the ESI source in positive ion mode. The transfer learning model is built on the PyTorch framework. The basic feature extraction layer contains four convolutional layers, each with 32 3×1 convolutional kernels. The input dimension is 200×1 (200 timestamps, each corresponding to a signal intensity), and the output is a 128-dimensional feature vector through the ReLU activation function. The MMD algorithm of the distribution adaptation module uses a Gaussian kernel function (bandwidth σ=10), and the loss function weights can be set to 0.38, 0.40, or 0.43, optimized through backpropagation. The output layer is a three-layer fully connected network with 128, 64, and 2 neurons respectively (corresponding to the start and end times of the window). During real-time acquisition of chromatographic signals, signals were collected at 0 seconds, 2 seconds, ..., 60 seconds after separation initiation, for a total of 20 seconds from the current time (i.e., 200 time points). The fragments were normalized to zero mean before being input into the model. For verification of the intensity of characteristic precursor ions in mass spectrometry, monitoring was performed using the mass spectrometer's PRM mode. The verification interval was 3 seconds before and after the start time, for a total of 6 seconds of data. The effective threshold was set to 1×10⁻⁶. 4 counts, invalid threshold is 5 × 10 3 Counts, incremental model training is performed in parallel on a multi-core CPU processor.

[0024] In existing technologies, traditional target fraction prediction methods often employ rule-based fixed window settings, which cannot adapt to retention time fluctuations of different samples. These methods frequently suffer from prediction window shifts due to minor changes in chromatographic conditions, leading to the omission of target fractions. This embodiment, however, combines a transfer learning model with the MMD algorithm to align the feature distributions of samples with different chemical properties. It dynamically updates the prediction model using real-time acquired chromatographic signals and employs a real-time verification mechanism based on the intensity of precursor ions from mass spectrometry. This significantly improves the prediction accuracy of the target fraction time window and reduces detection errors caused by prediction bias.

[0025] In another embodiment, incremental model training includes: adding real-time chromatographic signal fragments corresponding to marked invalid windows to the training dataset; freezing the parameters of the basic feature extraction layer; updating the output layer weights, with 15-20 training iterations; and using the updated transfer learning model to re-predict subsequent windows to generate the final target distillation time window set W1 to Wn.

[0026] For example, signal segments corresponding to invalid windows are stored in NumPy array format, containing timestamp sequences and normalized signal intensities, and are stored according to compound category when added to the training set. Parameter freezing of the basic feature extraction layer is achieved using PyTorch's `requires_grad=False` attribute, ensuring that the weights of the convolutional neural network are not updated during incremental training. The weights of the fully connected network in the output layer are updated using the RMSprop optimizer, with a learning rate set to 0.0015. The number of iterations can be selected as 16, 18, or 20, with each iteration using a batch size of 32 samples. After training, the model is re-inputted with real-time chromatographic data of the current separation stream, and the start and end times of subsequent windows are calculated through forward propagation with a timing accuracy of 0.1 seconds. The generated final window set is sent to the flow path control module via a message queue.

[0027] In existing technologies, when model predictions deviate, it is usually necessary to reload all training data for complete model training, which is time-consuming and affects the online detection process. However, this embodiment freezes the basic feature extraction layer and only performs incremental updates on the output layer. Combined with lightweight training of 15-20 iterations, model optimization can be completed within 10 seconds, avoiding the detection interruption caused by full training in traditional methods and ensuring the continuity and adaptability of the real-time detection process.

[0028] In another embodiment, when the first window P1 in the initial prediction window set is marked as invalid, the following operations are performed: reference chromatographic data with the same target compound category as the first test sample is retrieved from the historical database, and all target fraction time window parameters in the reference data are extracted as the default window set; before the transfer learning model completes incremental training, the default window set is overlaid on all windows to be verified in the initial prediction window set except for the first window; after the incremental training is completed, the default window set is immediately replaced with the newly generated final target fraction time window set.

[0029] For example, the historical database is indexed inverted by compound category, with category labels based on chemical structure and polarity. When reference data is invoked, a matching reference chromatographic dataset is retrieved by querying the compound category index. The reference data includes retention time axes, UV signal curves, and manually labeled time window parameters (start time, end time, peak width). The default window set's time window width can be 12 seconds, 15 seconds, or 18 seconds, with the parameters averaged from the three most similar reference datasets. Overlay operations are implemented using the `update` method of a Python dictionary, writing the start and end times of the default window to the corresponding positions in the prediction window set. The signal indicating incremental training completion is sent from the model training module to the control module via Socket communication, triggering window set replacement. The replacement process's time delay is controlled within 0.3 seconds to ensure the real-time nature of the flow path switching command.

[0030] In existing technologies, when the first prediction window is invalid, without an effective backup window strategy, the detection system may fail to determine the target fraction time period, leading to invalid data acquisition by the mass spectrometer or even interruption of the detection process. This embodiment, however, maintains the predictive ability of the target fraction during incremental model training by calling historical reference windows of similar compounds as the default set, avoiding detection interruptions due to prediction failure and ensuring the robustness of the method when analyzing complex samples.

[0031] In another embodiment, the determination of the same target compound category is achieved through the following steps: obtaining the precise molecular weight data, measured lipid-water partition coefficient data, and topological polar surface area data of the target compound in the first test sample; inputting the aforementioned molecular weight data, lipid-water partition coefficient data, and topological polar surface area data into a pre-trained compound classification model to generate a sample feature vector; calculating the similarity value between the sample feature vector and the feature vectors of all reference compounds in the historical database; when there is a reference compound with a similarity value greater than or equal to 0.9, determining that the reference compound and the first test sample belong to the same target compound category, and selecting the reference compound data source with the highest similarity value.

[0032] For example, precise molecular weight data were obtained using a Thermo Fisher Q Exactive HF mass spectrometer with a resolution of 70,000 and a mass accuracy controlled within ±3 ppm. The measured lipid-water partition coefficient (ClogP) was determined using the shake-flask method, equilibrated in a phosphate buffer and n-octanol system at pH 7.4, and then quantitatively calculated by HPLC. The topological polar surface area (TPSA) was calculated using Dragon software based on the molecular structure SMILES string. The pre-trained compound classification model was an XGBoost model containing 200 decision trees, with an input dimension of 3 (molecular weight, ClogP, TPSA) and an output of a 50-dimensional feature vector. Similarity was calculated using the reciprocal of the Euclidean distance, and the feature vector of each reference compound in the historical database was stored in HDF5 format. When the similarity was ≥0.9 (e.g., 0.91, 0.93, or 0.95), the compound was considered to belong to the same class. If multiple matches were found, the data source with the highest similarity (e.g., the reference compound with a similarity of 0.97) was selected, and its corresponding chromatographic data included the average time window parameters from at least three repeated experiments.

[0033] In existing technologies, compound classification often relies on manual comparison of molecular structures or simple molecular weight range divisions, which is highly subjective and cannot quantify the degree of similarity, easily leading to biases in reference data retrieval. This embodiment, however, achieves automated and accurate matching of target compound categories by combining quantified feature vectors and similarity calculations with a machine learning classification model. This provides data support for the retrieval of default window sets, making time window predictions for similar compounds more reliable.

[0034] In another embodiment, S4 includes: real-time acquisition of UV absorption signal intensity data of the first separated stream; calculation of the absolute value of the continuous signal change rate, where the signal change rate is defined as the absolute value of the difference between the absorbance values ​​of adjacent sampling points divided by a 0.1-second time interval; activation of mass spectrometry verification when the absolute value of the signal change rate does not exceed 0.05 AU / s for 50 consecutive data points; the mass spectrometry verification performs the following operations: acquisition of mass spectrometry characteristic precursor ion intensity data within the interval from 2 seconds before to 2 seconds after the trigger time; confirmation that the non-target distillation segment has been entered if the characteristic precursor ion intensity is lower than 1% of the maximum intensity of the current chromatographic peak; and sending a start command to the second liquid chromatography separation column at the start time of the first non-target distillation segment that meets the conditions.

[0035] For example, the UV absorption signal is acquired using a Shimadzu SPD-M30A diode array detector at a sampling frequency of 10 Hz, meaning the absorbance value at 254 nm is recorded every 0.1 seconds, with a dynamic range of 0-3 AU. The rate of change is calculated by subtracting the absorbance values ​​at the i-th and (i+1)-th sampling points, taking the absolute value, and dividing by 0.1 seconds to obtain the rate of change in AU / s. Mass spectrometry verification is triggered when 50 consecutive rate of change data (i.e., within 5 seconds) are all ≤0.05 AU / s. Mass spectrometry verification is performed using the AB Sciex Qtrap 6500+ mass spectrometer in MRM mode. Characteristic precursor ion intensities are acquired at 100 ms scan intervals within 2 seconds before and after the trigger (4 seconds in total). The maximum intensity of the current chromatographic peak is determined using a dynamic backtracking window. If the precursor ion intensity is less than 1% of the maximum intensity (e.g., the maximum intensity is 5 × 10⁻⁶), the verification is performed. 5 If counts, then the threshold is 5 × 10 3 (counts), confirming entry into a non-target section, the start command is sent to the pump system of the second separation column via the GPIB interface to control the start of the mobile phase gradient.

[0036] In existing technologies, determining non-target segments solely based on the rate of change of ultraviolet signals may lead to misjudgments due to slow baseline drift or noise interference, resulting in the incorrect activation of another channel. This embodiment, however, employs a dual verification mechanism combining continuous monitoring of the signal change rate with the intensity of characteristic precursor ions in mass spectrometry. Non-target segments are only confirmed when both conditions are met simultaneously, effectively reducing the probability of false activation, making the activation timing of the second separation column more precise, and avoiding redundant channel switching operations.

[0037] In another embodiment, determining the maximum intensity of the current chromatographic peak in S4.4 includes: using the moment of activation of mass spectrometry verification as the reference time trigger; setting a dynamic backtracking time interval [tstart, ttrigger], where the start time tstart is calculated by the following formula: tstart=ttrigger−min(120, 3×Wavg), where Wavg is the average chromatographic peak width of the target compound in the current liquid chromatography system, and its value is obtained through calibration experiments during the system initialization phase. The calibration method is as follows: performing chromatographic separation on at least 5 standard target compounds and calculating the average half-peak width of the main chromatographic peak of each compound; extracting the maximum intensity of the ultraviolet absorption signal within the time interval [tstart, ttrigger], denoted as Imax, and outputting Imax as the maximum intensity of the current chromatographic peak.

[0038] For example, the base time trigger is accurate to 0.1 seconds by the system clock and stored in Unix timestamp format. The t of the dynamically backtracking time interval... startIn the calculation, Wavg is obtained through a calibration experiment. For example, five standard substances, namely benzoic acid, aniline, o-xylene, phenanthrene, and cholesterol, are selected. Chromatographic separation is performed with an injection volume of 10 μL for each. The Chromeleon software is used to automatically identify the full width at half maximum (FWHM) of the main chromatographic peaks of each compound, with the unit being seconds. The average value is calculated to obtain Wavg (for example, if Wavg = 30 seconds, then 3×Wavg = 90 seconds, and tstart = trigger - 90 seconds; if Wavg = 50 seconds, then 3×Wavg = 150 seconds, take min(120, 150) = 120 seconds, and tstart = trigger - 120 seconds). The ultraviolet signal intensity data is stored in a circular buffer with a size of 2000 time points and a time resolution of 0.1 seconds. By traversing the buffer data within the interval [tstart, ttrigger], the numpy.max() function is used to extract the maximum value Imax, with the unit being AU, which is used for the subsequent calculation of the precursor ion intensity threshold (such as 1% of Imax).

[0039] In the prior art, when determining the maximum intensity of a chromatographic peak using a fixed backtracking time (such as a fixed 120 seconds), if the current chromatographic peak width is small, the backtracking interval may include multiple irrelevant peaks, resulting in an incorrect extraction of the maximum intensity; if the peak width is large, the true peak may be missed due to insufficient backtracking interval. In this embodiment, through the dynamic backtracking time formula, the backtracking interval is adaptively adjusted according to the average peak width calibrated by the system, ensuring that the maximum intensity of the current peak can be accurately captured under different chromatographic conditions, providing a reliable threshold benchmark for mass spectrometry verification, and improving the accuracy of non-target segment determination.

[0040] In another embodiment, S6 further includes: when the current time t belongs to any window Wi in the target fraction time window set of the first separation flow and any window Vj in the target fraction time window set of the second separation flow at the same time, perform the following operations: compare the start time stamp T1 of the first liquid chromatography separation column with the start time stamp T2 of the second liquid chromatography separation column; if T1 < T2, introduce the first separation flow into the mass spectrometer and switch the second separation flow to the waste liquid channel at the same time; if T2 < T1, introduce the second separation flow into the mass spectrometer and switch the first separation flow to the waste liquid channel at the same time.

[0041] For example, timestamps T1 and T2 are generated by the PLC controller's real-time clock with an accuracy of 10ms and written to the register via the Modbus protocol when the separation column starts. When overlapping time windows are detected, the control software reads and compares the values ​​of T1 and T2. Flow path switching is achieved through a two-position three-way solenoid valve (such as the Lee Valve 173 series), with a switching delay time ≤30ms. The pipes for the first and second separation streams are connected to the two inlet terminals of the solenoid valve, respectively. The mass spectrometer inlet is connected to the common terminal, and the waste liquid channel is connected to the other outlet terminal. The switching logic is as follows: if T1 is earlier than T2 (e.g., T1=14:30:25.000, T2=14:30:28.000), it indicates that the first separation column starts first, and its separation stream is preferentially introduced into the mass spectrometer. The second separation stream is temporarily discharged into the waste liquid to avoid overloading the mass spectrometer ion source or data overlap due to simultaneous sample injection.

[0042] In existing technologies, when the target fraction time windows of dual channels overlap, the lack of a clear priority control strategy may lead to both separation streams being connected to the mass spectrometer simultaneously, causing ion suppression or data corruption and affecting the reliability of the detection results. This embodiment, however, establishes a "first-in, first-out" principle by comparing the sequence of separation column start timestamps, ensuring that the mass spectrometer processes only one separation stream during the overlapping period. This avoids the impact of channel conflicts on detection quality and guarantees the accuracy and integrity of the mass spectrometry data.

[0043] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and embodiments shown and described herein.

Claims

1. A real-time modulated multiplexed tandem mass spectrometry detection method, characterized in that, include: S1: The pre-treated first sample to be tested is loaded into the first injection channel, and the pre-treated second sample to be tested is loaded into the second injection channel. The first injection channel and the second injection channel are physically isolated. S2: Start the first liquid chromatography column to separate the first sample to be tested, generating the first separation stream; S3: Based on the real-time chromatographic data of the first separated stream, dynamically predict its target fraction time window set using a machine learning model. The target fraction time window set includes n consecutive time periods W1 to W... n W for each time period i From start time t i -start and end times t i -end definition; S4: Monitor the chromatographic signal change rate of the first separation stream in real time. When the absolute value of the signal change rate does not exceed 0.05 AU / s for 5 consecutive seconds, it is determined that the stream has entered a non-target distillation segment. At the beginning of the first non-target distillation segment, the second liquid chromatography separation column is started to process the second sample to be tested and generate the second separation stream. S5: Perform the dynamic prediction of S3 on the second separated stream to generate its target fraction time window set, which contains m consecutive time periods V1 to V... m V for each time period j From start time t j -start and end times t j -end definition; S6: If the current time t belongs to any time period Wᵢ in the target fraction time window set of the first separation stream, the first separation stream is fed into the mass spectrometer; if the current time t belongs to any time period Vⱼ in the target fraction time window set of the second separation stream, the second separation stream is fed into the mass spectrometer; if the current time t does not belong to any target fraction time window, the output stream is switched to the waste liquid channel.

2. The method as described in claim 1, characterized in that, S3 include: S3.1: Obtain 200 to 300 sets of chromatographic-mass spectrometry data of test samples with different chemical properties from the historical detection database. Each set of data includes the retention time axis, ultraviolet absorption signal intensity and mass spectrometry characteristic precursor ion intensity. S3.2: Construct a transfer learning model framework, which includes a basic feature extraction layer, a distribution adaptation module, and an output layer. The basic feature extraction layer uses a convolutional neural network to process chromatographic signals, with the input dimension being the timestamp multiplied by the signal intensity, and the output being a 128-dimensional feature vector. The distribution adaptation module aligns the feature distributions of different test objects using the maximum mean difference algorithm, with the maximum mean difference loss function weight set to 0.35-0.

45. The output layer is a fully connected neural network structure. S3.3: For the first sample to be tested, within 0 to 60 seconds after separation starts, acquire a real-time chromatographic signal fragment every 2 seconds for a duration of 20 seconds; input the real-time chromatographic signal fragment into the transfer learning model in S3.2, and output the initial prediction window set P1 to P... k ; S3.4: When the retention time of the first separated stream reaches the start time t of the first window P1 in the initial prediction window set. p1 When -start is executed, the intensity verification of the characteristic precursor ion of mass spectrometry is performed: In t p1 -start 3 seconds before t p1 If the intensity of the characteristic precursor ion of the target compound exceeds 1×10⁻⁶ within 3 seconds after starting, 4 If counts are found, then P1 is confirmed to be valid. If the characteristic precursor ion strength does not exceed 5 × 10 3 If counts are found, then P1 is marked as invalid and incremental training of the model is activated.

3. The method as described in claim 2, characterized in that, The incremental training of the model includes: adding real-time chromatographic signal segments corresponding to marked invalid windows to the training dataset; freezing the parameters of the basic feature extraction layer; updating the output layer weights, with 15-20 training iterations; The updated transfer learning model is used to re-predict subsequent windows, generating the final target fraction time window set W1 to W2. n .

4. The method as described in claim 3, characterized in that, When the first window P1 in the initial prediction window set is marked as invalid, perform the following operations: Retrieve reference chromatographic data with the same target compound class as the first test sample from the historical database, and extract all target fraction time window parameters from the reference data as the default window set; Before the transfer learning model completes incremental training, the default window set is overlaid on all the windows to be verified in the initial prediction window set except for the first window; Once incremental training is complete, the default window set is immediately replaced with the newly generated final target distillation time window set.

5. The method as described in claim 4, characterized in that, The determination of the same target compound category is achieved through the following steps: Obtain accurate molecular weight data, measured lipid-water partition coefficient data, and topological polar surface area data of the target compound in the first test sample; The aforementioned molecular weight data, lipid-water partition coefficient data, and topological polar surface area data are input into a pre-trained compound classification model to generate sample feature vectors. Calculate the similarity between the feature vector of this sample and the feature vectors of all reference compounds in the historical database; When there is a reference compound with a similarity value greater than or equal to 0.9, it is determined that the reference compound and the first sample to be tested belong to the same target compound category, and the data source of the reference compound with the highest similarity value is selected.

6. The method as described in claim 1, characterized in that, S4 includes: S4.1: Real-time collect the ultraviolet absorption signal intensity data of the first separated flow; S4.2: Calculate the absolute value of the continuous signal change rate, where the signal change rate is defined as the absolute value of the difference between the absorbance values of adjacent sampling points divided by a 0.1-second time interval; S4.3: When it is detected that the absolute value of the signal change rate of 50 consecutive data points does not exceed 0.05 AU / s, activate the mass spectrometry verification; S4.4: The mass spectrometry verification performs the following operations: Collect the mass spectrometry characteristic parent ion intensity data within the interval from 2 seconds before the trigger moment to 2 seconds after; If the intensity of the characteristic parent ion is lower than 1% of the maximum intensity of the current chromatographic peak, it is confirmed that the non-target fraction section is entered; S4.5: At the starting moment of the first non-target fraction section that meets the conditions of S4.4, send a start command to the second liquid chromatography separation column.

7. The method as described in claim 6, characterized in that, The determination of the maximum intensity of the current chromatographic peak in S4.4 includes: Taking the moment when the mass spectrometry verification is activated as the reference moment trigger; Set the dynamic backtracking time interval [t] start , t trigger ], where the start time t start Calculate using the following formula: t start =t trigger -min(120, 3×W avg ) In the formula W avg The average peak width of the target compound in the current liquid chromatography system is obtained through calibration experiments during the system initialization phase. The calibration method is as follows: perform chromatographic separation on at least 5 standard target compounds and calculate the average half-peak width of the main chromatographic peak of each compound. Extract time interval [t] start , t trigger The maximum ultraviolet absorption signal intensity within [the range] is denoted as I. max , will I max Output as the maximum intensity of the current chromatographic peak.

8. The method as described in claim 1, characterized in that, S6 also includes: When the current time t simultaneously belongs to any window W in the target fraction time window set of the first separation stream i Any window V in the target fraction time window set of the second separated stream j When this happens, perform the following operations: Compare the start timestamp T1 of the first liquid chromatography separation column with the start timestamp T2 of the second liquid chromatography separation column; If T1 < T2, introduce the first separated flow into the mass spectrometer and switch the second separated flow to the waste liquid channel at the same time; If T2 < T1, introduce the second separated flow into the mass spectrometer and switch the first separated flow to the waste liquid channel at the same time.

Images