Method for rapid detection of four immunosuppressant concentrations simultaneously

Abstract

The application discloses a method for simultaneously and rapidly detecting the concentration of four immunosuppressants, comprising the following steps: mixing a blood sample with an internal standard solution containing ascomycin, cyclosporine-d4, sirolimus-d3 and everolimus-d4, adding a precipitant and centrifuging to obtain a sample to be detected; injecting into a liquid chromatography-tandem mass spectrometry system for analysis to obtain original data containing the retention time window and mass spectrum response signal of four analytes; searching the original data based on an interferent database, wherein the interferent database contains a plurality of compounds that can possibly generate mass spectrum signal interference and characteristic ion information thereof; inputting the original data and the searched potential interferent information into a machine learning model to output a net mass spectrum response value after interference correction; and calculating the concentration of tacrolimus, cyclosporine, sirolimus and everolimus in the blood sample according to the obtained net mass spectrum response value, the mass spectrum response value of the corresponding internal standard and a pre-established standard curve.

Description

Technical Field

[0001] This invention relates to the field of medical testing technology. More specifically, this invention relates to a method for the simultaneous and rapid detection of the concentrations of four immunosuppressants. Background Technology

[0002] In post-organ transplant treatment, tacrolimus, cyclosporine, sirolimus, and everolimus are commonly used immunosuppressants. Monitoring their concentrations is crucial for ensuring treatment efficacy and reducing the risk of adverse reactions. These drugs have a narrow therapeutic window and exhibit significant individual metabolic variability. Excessively high blood concentrations can easily lead to serious adverse reactions such as nephrotoxicity and neurotoxicity, while excessively low concentrations may cause transplant rejection. Therefore, accurate and rapid quantitative detection of blood drug concentrations is essential.

[0003] Currently, commonly used clinical methods for detecting immunosuppressants mainly include immunoassay and liquid chromatography-tandem mass spectrometry. Immunoassay is simple to operate and has a relatively fast detection speed, but it has poor specificity and is easily affected by cross-interference from endogenous substances, other coexisting drugs, and their metabolites in the sample, leading to biased test results. It is especially unsuitable for the accurate simultaneous detection of multiple components.

[0004] Liquid chromatography-tandem mass spectrometry (LC-MS / MS) has gradually become the mainstream method for immunosuppressant detection due to its high sensitivity and specificity. However, existing detection schemes based on this technology still have many limitations. Interfering substances such as endogenous metabolites and exogenous coexisting drugs in blood samples can easily generate superposition of mass spectrometric signals within the retention time window of the target analyte, leading to distortion of the target analyte response value and affecting the accuracy of the detection results.

[0005] Therefore, it is necessary to design a technical solution that can overcome the above-mentioned defects. Summary of the Invention

[0006] One object of the present invention is to provide a method for simultaneously and rapidly detecting the concentrations of four immunosuppressants, which can improve the detection accuracy and specificity of the four immunosuppressants.

[0007] To achieve these objectives and other advantages of the present invention, according to one aspect of the present invention, a method for the simultaneous and rapid detection of four immunosuppressant concentrations is provided, comprising: S1: mixing a blood sample with an internal standard solution containing ascomycin, cyclosporine-d4, sirolimus-d3, and everolimus-d4, adding a precipitant, and centrifuging to obtain a sample to be tested; S2: injecting the sample to be tested into a liquid chromatography-tandem mass spectrometry system for analysis, obtaining raw data of retention time windows and mass spectrometric response signals of the four analytes, tacrolimus, cyclosporine, sirolimus, and everolimus; S3: analyzing the sample from S2 based on a pre-constructed database of interfering substances. The raw data is retrieved, and the interfering substance database contains information on various compounds and their characteristic ions that may cause mass spectrometry signal interference within the retention time windows of the four analytes; S4: The raw data obtained in S2 and the potential interfering substance information retrieved in S3 are input into a pre-trained machine learning model, which outputs the interference-corrected net mass spectrometry response values ​​for tacrolimus, cyclosporine, sirolimus, and everolimus in the current sample; S5: Based on the net mass spectrometry response value obtained in step S4, the mass spectrometry response value of the corresponding internal standard, and the pre-established standard curve, the concentrations of tacrolimus, cyclosporine, sirolimus, and everolimus in the blood sample are calculated.

[0008] Further, in S2, the sample to be tested is injected into a liquid chromatography system equipped with a C18 reversed-phase column for separation. Mobile phase A is an aqueous solution containing 0.1% formic acid and 2mM ammonium acetate, and mobile phase B is a methanol solution containing 0.1% formic acid, using a gradient elution program. The effluent after chromatographic separation enters a tandem mass spectrometry system and is detected using electrospray positive ionization mode and multiple reaction monitoring mode.

[0009] Furthermore, in S3, the compounds included in the interfering substance database include endogenous substances, exogenous drugs, and their metabolites that have been observed to interfere with mass spectrometry signals within the specific retention time windows of the four analytes. Each compound is recorded with its name, chemical formula, typical retention time range, at least one pair of characteristic precursor ion and daughter ion pairs, and their relative response intensity benchmark information under the detection system. During the search, the ion pair signals of the original data obtained in S2, which are preset to be monitored for the interfering substance database within the retention time windows of each analyte, are matched with the characteristic ion pairs of interfering substances stored in the interfering substance database. For suspected interfering substances that are successfully matched, the intensity of their mass spectrometry response signal in the current sample is further compared with the relative response intensity benchmark of the interfering substance recorded in the database under the same conditions. If its response intensity exceeds the preset interference judgment threshold, the compound is marked as a valid potential interfering substance.

[0010] Furthermore, in S4, the pre-trained machine learning model is a supervised learning model, and the input data received by the input layer includes: (a) the original quantitative ion pair response signal obtained in S2 within their respective retention time windows for each of the four analytes: tacrolimus, cyclosporine, sirolimus, and everolimus; (b) the quantitative ion pair response signal obtained within the same retention time window for the internal standards corresponding to the four analytes; (c) the characteristic ion pair response signal monitored within the corresponding analyte retention time window for each compound marked as an effective potential interfering agent in S3; and (d) the response intensity ratio of each effective potential interfering agent to the corresponding analyte internal standard calculated based on the signals in (b) and (c).

[0011] Furthermore, the machine learning model includes a first path, a second path, and a cross-attention module. The first path contains at least one fully connected layer for extracting features of the standardized analyte response values ​​and internal standard response values ​​to generate a first feature vector. The second path contains at least one fully connected layer for extracting features of the standardized response values ​​of each effective potential interfering analyte and their response intensity ratios with the corresponding analyte internal standard to generate a second feature vector. The cross-attention module uses the first feature vector generated by the first path as the query vector and the second feature vector generated by the second path as the key vector and value vector. By calculating the attention weight between the query vector and the key vector, the value vector is weighted and summed to generate a context-aware interference feature vector. The context-aware interference feature vector is concatenated with the first feature vector and input together into the subsequent fully connected layer for calculation. The subsequent fully connected layer outputs a four-dimensional interference-corrected net mass spectrometry response value vector and a four-dimensional interference correction factor vector. The interference correction factor vector represents the correction multipliers calculated by the model and applied to the original response values ​​of the four analytes.

[0012] Furthermore, the training method for the machine learning model includes: constructing a training dataset, which includes a synthetic spiked training set and a real clinical validation set; the synthetic spiked training set is obtained by adding different known concentrations of tacrolimus, cyclosporine, sirolimus, and everolimus standards to a blank matrix to form a set of basic calibration samples; in the basic calibration samples, multiple interfering substance standards selected from an interfering substance database are further added, with the types and concentrations of interfering substances designed based on their clinically common ranges to form multiple sets of spiked interfering samples; the real clinical validation set is obtained by collecting clinical samples that have undergone quantitative detection using a validated reference method; defining a composite loss function L, whose expression is: L = L MSE +λ*L REG； Among them, L MSE L is the mean square error between the interference-corrected net mass spectrum response vector output by the model and the corresponding training label;REG The L2 norm regularization penalty term is applied to the interference correction factor vector of the model's synchronous output, and λ is the regularization strength coefficient. During training, the machine learning model is first trained using a synthetic spiked training set, and the composite loss function L is minimized through an optimization algorithm. Then, the performance of the trained model is evaluated using a real clinical validation set.

[0013] Further, S5 includes: S51: Read the interference correction factor vector output by the machine learning model in S4. If the values ​​of all four elements are within the preset reference range, proceed to S52; if the value of any element exceeds the reference range, proceed to S53; S52: For tacrolimus, cyclosporine, sirolimus, and everolimus, divide the net mass spectrometry response value obtained in S4 by the mass spectrometry response value of the corresponding internal standard to obtain the internal standard normalized response value; substitute the internal standard normalized response value of each analyte into the pre-established corresponding polynomial standard curve to calculate its concentration; S53: For dry... For analytes whose values ​​in the interference correction factor vector exceed the baseline range, dynamic curve correction is performed. Specifically: S531: The pre-stored standard curve equation for the analyte is called, which is a polynomial standard curve; S532: Using the interference correction factor of the analyte output by the machine learning model as a dynamic parameter, the function shape of the standard curve equation is adjusted for adaptability through a predefined curve correction function to generate an instantaneous correction curve that is only used for the current sample; S533: The ratio of the net mass spectrometry response value of the analyte to the internal standard response value is substituted into the instantaneous correction curve to calculate its concentration.

[0014] Further, based on the compound category information of the effective potential interfering substances marked in S3, the type of interference is determined; if the effective potential interfering substances are mainly endogenous substances, the first correction function is called to correct the polynomial standard curve of the analyte, and the first correction function is configured to perform nonlinear correction for the low concentration segment of the curve; if the effective potential interfering substances are mainly exogenous drugs or metabolites, the second correction function is called to correct the polynomial standard curve, and the second correction function is configured to perform nonlinear correction for the high concentration segment of the curve; the interference correction factor of the analyte output by the machine learning model is used as the input parameter and substituted into the called correction function to generate an instantaneous correction curve that is only used for the current sample.

[0015] The present invention has at least the following beneficial effects: This invention accurately retrieves potential interfering substances by constructing an interfering substance database and uses a machine learning model to correct for interference, effectively eliminating interference from endogenous substances, exogenous drugs, and their metabolites, significantly improving detection accuracy and specificity. The overall method is simple to operate, with a concise pretreatment process that eliminates the need for complex separation steps, balancing practicality and accuracy. It can be widely applied to routine clinical testing, effectively solving the problems of existing methods, such as difficulty in simultaneously and accurately detecting multiple immunosuppressants, susceptibility to interference, and low detection efficiency.

[0016] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0017] Figure 1 This is a flowchart of one embodiment of this application. Detailed Implementation

[0018] The present invention will now be described in further detail so that those skilled in the art can implement it based on the description.

[0019] 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.

[0020] 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.

[0021] like Figure 1As shown, embodiments of this application provide a method for the simultaneous and rapid detection of four immunosuppressant concentrations, comprising: S1: mixing a blood sample with an internal standard solution containing ascomycin, cyclosporine-d4, sirolimus-d3, and everolimus-d4, adding a precipitant, and centrifuging to obtain the sample to be tested; S2: injecting the sample to be tested into a liquid chromatography-tandem mass spectrometry system for analysis, obtaining raw data of the retention time windows and mass spectrometry response signals of the four analytes, tacrolimus, cyclosporine, sirolimus, and everolimus; S3: retrieving the raw data from S2 based on a pre-constructed database of interfering substances, and filtering out the interfering substances. The material database contains information on various compounds and their characteristic ions that are known to cause mass spectrometry signal interference within the retention time windows of the four analytes; S4: Input the raw data obtained in S2 and the potential interfering information retrieved in S3 into a pre-trained machine learning model, and output the interference-corrected net mass spectrometry response values ​​for tacrolimus, cyclosporine, sirolimus, and everolimus in the current sample; S5: Calculate the concentrations of tacrolimus, cyclosporine, sirolimus, and everolimus in the blood sample based on the net mass spectrometry response value obtained in step S4, the corresponding internal standard mass spectrometry response value, and the pre-established standard curve.

[0022] For example, a blood sample is mixed with an internal standard solution containing ascomycin, cyclosporine-d4, sirolimus-d3, and everolimus-d4. The blood sample can be a venous whole blood sample from an organ transplant patient. Each internal standard in the solution is a stable isotope internal standard or a recognized internal standard (similar molecular weight, consistent peak time) for the corresponding analyte. Its function is to correct for losses during sample pretreatment and matrix effects during instrument detection. The concentration of each internal standard in the solution can be 5 ng / mL or 10 ng / mL. Subsequently, a precipitant is added and centrifuged to obtain the sample to be tested. The precipitant can be methanol or acetonitrile, and the amount added can be twice the volume of the blood sample. After addition, the sample is vortexed at 2000 r / min for 1 minute and then centrifuged at 8000 r / min for 5 minutes. The purpose of this step is to denature and precipitate the proteins in the blood sample, thereby releasing the target analyte into the solution and separating it from the protein. The test samples were injected into a liquid chromatography-tandem mass spectrometry (LC-MS / MS) system for analysis to obtain raw data on the retention time windows and mass spectrometric response signals of four analytes: tacrolimus, cyclosporine, sirolimus, and everolimus. The LC-MS / MS system is an analytical device integrating chromatographic separation and mass spectrometry detection, used to achieve the separation and accurate detection of the four analytes. The raw data includes information such as the ion signal intensity of each analyte within a specific retention time range. The raw data was then retrieved based on a pre-constructed interference substance database. This database, which pre-collects, organizes, and stores information on various interfering compounds, is used to quickly identify potential interference signals in the raw data. The database contains information on various compounds known to potentially interfere with mass spectrometric signals within the retention time windows of the four analytes, along with their characteristic ion information, including the parent ion mass-to-charge ratio and the daughter ion mass-to-charge ratio. The acquired raw data and the retrieved information on potential interfering substances are input into a pre-trained machine learning model. The model outputs the interference-corrected net mass spectrometry response values ​​for the four analytes in the current sample. This machine learning model, trained on a large number of samples, is a computational model capable of interference identification and correction. Its function is to eliminate the influence of interfering substances on the analyte mass spectrometry response signal, ensuring the accuracy of the response values. Based on the obtained net mass spectrometry response values, the corresponding internal standard mass spectrometry response values, and a pre-established standard curve, the concentrations of the four analytes in the blood sample are calculated. This step converts the corrected mass spectrometry response values ​​into specific concentration values. The calculation process is as follows: first, the ratio of the analyte's net mass spectrometry response value to the corresponding internal standard mass spectrometry response value is calculated; then, this ratio is substituted into the standard curve equation to obtain the analyte concentration. The standard curve equation can be y=ax+b or y=ax. 2 +bx+c, where y is the ratio of the analyte's response to the internal standard, x is the analyte concentration, and a, b, and c are the fitting parameters of the standard curve.

[0023] This embodiment can detect four analytes simultaneously without the need for multiple separate tests, significantly improving detection efficiency. At the same time, by combining an interfering substance database with a machine learning model, it effectively identifies and corrects interference signals. Compared with existing technologies, the detection results are more accurate and meet the needs of clinical combined drug monitoring. Its method of combining database retrieval with machine learning to achieve interference correction solves the problem of difficulty in eliminating interference when detecting multiple components in existing technologies.

[0024] In another embodiment, in S2, the sample to be tested is injected into a liquid chromatography system equipped with a C18 reversed-phase column for separation. Mobile phase A is an aqueous solution containing 0.1% formic acid and 2mM ammonium acetate, and mobile phase B is a methanol solution containing 0.1% formic acid, using a gradient elution program. The effluent after chromatographic separation enters a tandem mass spectrometry system and is detected using electrospray positive ionization mode and multiple reaction monitoring mode.

[0025] For example, in the step of injecting the sample to be analyzed into a liquid chromatography-tandem mass spectrometry system, the sample is injected into a liquid chromatography system equipped with a C18 reversed-phase column for separation. The C18 reversed-phase column is a chromatographic separation column with octadecylsilane as the stationary phase, and its function is to achieve separation by utilizing the difference in the partition coefficients of different analytes between the stationary phase and the mobile phase. The specifications of the chromatographic column can be DISIGNS. Column-002 (30mm*3.0mm); Mobile phase A is an aqueous solution containing 0.1% formic acid and 2mM ammonium acetate, and mobile phase B is a methanol solution containing 0.1% formic acid. The role of formic acid is to improve the ionization efficiency of the analyte, and ammonium acetate can enhance the buffering capacity of the mobile phase. 0.1% is the volume percentage of formic acid in the mobile phase, and 2mM is the molar concentration of ammonium acetate. A gradient elution program is used. The purpose of the gradient elution program is to change the volume ratio of mobile phase A and mobile phase B so that analytes with different retention characteristics can be eluted and separated at appropriate times. See the table below for details: 0-0.5 min, mobile phase A volume fraction 85%; 0.5-1.5 min, mobile phase A volume fraction 2%; 1.5-2.8 min, mobile phase A volume fraction 85%. The effluent after chromatographic separation enters a tandem mass spectrometry system and is detected using electrospray positive ionization mode and multiple reaction monitoring mode. Electrospray positive ionization mode is a method of ionization that converts analyte molecules into positively charged ions, while multiple reaction monitoring mode is a mode that selects specific precursor ions and daughter ions for detection, which improves the sensitivity and specificity of detection. The ion source temperature can be 350℃ and the spray voltage can be 3500V.

[0026] This embodiment employs a mobile phase with a specific composition and an optimized gradient elution program, combined with a C18 reversed-phase column, to achieve efficient separation of four analytes. At the same time, specific ionization and monitoring modes improve ionization efficiency and detection specificity. Compared with existing technologies, the separation effect is better and the detection sensitivity is higher, solving the problems of incomplete separation of multiple components and poor detection stability in existing technologies.

[0027] In another embodiment, in S3, the compounds included in the interfering substance database include endogenous substances, exogenous drugs, and their metabolites that have been observed to interfere with the mass spectrometry signal within the specific retention time windows of the four analytes. Each compound is recorded with its name, chemical formula, typical retention time range, at least one pair of characteristic precursor ion and daughter ion pairs, and their relative response intensity benchmark information under the detection system. During the search, the ion pair signals of the interfering substance database that are preset to be monitored in the raw data obtained in S2 within the retention time windows of each analyte are matched with the interfering substance characteristic ion pairs stored in the interfering substance database. For suspected interfering substances that are successfully matched, the intensity of their mass spectrometry response signal in the current sample is further compared with the relative response intensity benchmark of the interfering substance under the same conditions recorded in the database. If its response intensity exceeds the preset interference judgment threshold, the compound is marked as a valid potential interfering substance.

[0028] For example, in the step of retrieving raw data based on a pre-built interfering substance database, the compounds included in the interfering substance database include endogenous substances, exogenous drugs and their metabolites that have been observed to interfere with the mass spectrometry signal within four specific retention time windows of the analytes. The endogenous substances can be bilirubin or cholesterol, the exogenous drugs can be antibiotics or antifungal drugs, and the metabolites can be hydroxylated or demethylated metabolites of the drugs. The role of this database is to provide basic data support for the identification of interfering substances. Each compound is recorded with its name, chemical formula, typical retention time range, at least one pair of characteristic precursor ion and daughter ion pairs and their relative response intensity reference information under the detection system. The characteristic precursor ion and daughter ion pairs can be the m / z of tacrolimus interfering substances, and the relative response intensity reference information is the reference value of the response intensity of the compound under standard detection conditions. During the search, the ion pair signals of each analyte within the retention time window of the original data are first extracted for monitoring against the interference substance database. Then, these signals are matched with the characteristic ion pairs of interference substances stored in the database. The matching method is to compare whether the mass-to-charge ratio difference of the ion pairs is within the allowable error range, which can be ±0.1. For suspected interference substances that are successfully matched, the intensity of their mass spectrometry response signal in the current sample is further compared with the relative response intensity benchmark of the interference substance under the same conditions recorded in the database. If the response intensity exceeds the preset interference judgment threshold, the compound is marked as a valid potential interference substance. The interference judgment threshold can be 1.2 times the relative response intensity benchmark. The purpose of this step is to accurately screen out interference substances that have a real impact on the detection of analytes.

[0029] In existing technologies, the detection of immunosuppressants often lacks a dedicated interfering substance search, relying solely on chromatographic separation to reduce interference. This makes it difficult to identify co-eluted interfering substances, resulting in significant interference-induced impact on detection results. This embodiment constructs a database containing detailed information on various interfering substances and uses specific matching and comparison logic to screen for effective interfering substances. It can accurately identify interfering substances, such as those that are difficult to eliminate through chromatographic separation, providing an accurate basis for subsequent interference correction. Compared to existing technologies, interference identification is more comprehensive and accurate, solving the problem of insufficient interfering substance identification in existing technologies.

[0030] In another embodiment, in S4, the pre-trained machine learning model is a supervised learning model, and the input data received by the input layer includes: (a) the original quantitative ion pair response signal obtained in S2 within their respective retention time windows for each of the four analytes: tacrolimus, cyclosporine, sirolimus, and everolimus; (b) the quantitative ion pair response signal obtained within the same retention time window for the internal standards corresponding to the four analytes; (c) the characteristic ion pair response signal monitored within the corresponding analyte retention time window for each compound marked as a potential interfering agent in S3; and (d) the response intensity ratio of each potential interfering agent to the corresponding analyte internal standard calculated based on the signals in (b) and (c).

[0031] For example, in the step of inputting raw data and information on potential interfering substances into a pre-trained machine learning model to output the net mass spectrometry response value after interference correction, the pre-trained machine learning model is a supervised learning model. A supervised learning model learns the mapping relationship between input and output using labeled training data, and its function is to use the learned rules to achieve interference correction. The input data received by the input layer includes four categories: the first category is the raw quantitative ion pair response signal obtained by each of the four analytes within its respective retention time window; the raw quantitative ion pair response signal is the characteristic ion pair signal intensity used for quantification of the analyte in mass spectrometry detection; the second category is the quantitative ion pair response signal obtained by the internal standard corresponding to the four analytes within the same retention time window; the internal standard quantitative ion pair can be the m / z of ascosin. 809.6→757.6; The third category is the characteristic ion pair response signal detected for each compound marked as an effective potential interfering agent within the corresponding analyte retention time window; the fourth category is the ratio of the response intensity of each effective potential interfering agent to the corresponding analyte internal standard, calculated based on the signals from the second and third categories. The calculation process involves dividing the characteristic ion pair response signal intensity of each effective potential interfering agent by the quantitative ion pair response signal intensity of the corresponding analyte internal standard. The purpose of these four categories of input data is to provide the model with comprehensive detection signals and interference-related information to ensure the accuracy of correction.

[0032] In existing technologies, even when interfering objects are identified, simple signal subtraction methods are often used for interference correction, without fully considering the complex relationships between the interfering objects, the analyte, and the internal standard, resulting in poor correction performance. This embodiment clarifies the composition of the input data for the supervised learning model, especially including the response intensity ratio between the interfering object and the internal standard. This allows the model to fully utilize the correlation patterns between various signals for correction. Compared to existing technologies, the correction logic is more comprehensive, the correction effect is better, and it solves the problem of simple and ineffective interference correction methods in existing technologies.

[0033] In another embodiment, the machine learning model includes a first path, a second path, and a cross-attention module. The first path contains at least one fully connected layer for extracting features of the standardized analyte response values ​​and internal standard response values ​​to generate a first feature vector. The second path contains at least one fully connected layer for extracting features of the standardized response values ​​of each effective potential interfering analyte and their response intensity ratios with the corresponding analyte internal standard to generate a second feature vector. The cross-attention module uses the first feature vector generated by the first path as a query vector and the second feature vector generated by the second path as a key vector and a value vector. By calculating the attention weight between the query vector and the key vector, the value vector is weighted and summed to generate a context-aware interference feature vector. The context-aware interference feature vector is concatenated with the first feature vector and input together into the subsequent fully connected layer for calculation. The subsequent fully connected layer outputs a four-dimensional interference-corrected net mass spectrometry response value vector and a four-dimensional interference correction factor vector. The interference correction factor vector represents the correction multipliers calculated by the model and applied to the original response values ​​of the four analytes.

[0034] For example, the machine learning model includes a first path, a second path, and a cross-attention module. The function of this model structure is to extract features related to the analyte and the interfering substance, respectively, and establish the correlation between them to achieve accurate interference correction. The first path contains at least one fully connected layer, which is a neural network layer that fully connects all neurons in the previous layer to all neurons in the current layer. Its function is to extract and transform features from the input data. This path is used to extract features of the standardized analyte response value and the internal standard response value and generate a first feature vector. The standardization process involves dividing the response value by a preset standard response value benchmark, which can be 1000. The first feature vector is vector data containing the core features of the analyte and the internal standard. The second path contains at least one fully connected layer, used to extract features of the standardized response values ​​of each effective potential interfering substance and their ratio to the response intensity of the corresponding analyte and internal standard, generating a second feature vector. Here, the standardization process is the same as in the first path. The second feature vector is vector data containing the core features of the interfering substance. The cross-attention module uses the first feature vector generated by the first path as the query vector and the second feature vector generated by the second path as the key and value vectors. By calculating the attention weights between the query vector and the key vector, the value vectors are weighted and summed to generate a context-aware interference feature vector. The attention weights can be calculated using the Softmax function, which involves first calculating the dot product of the query vector and the key vector, then dividing by the square root of the key vector dimension, and finally normalizing using the Softmax function to obtain the weights. The context-aware interference feature vector is concatenated with the first feature vector and fed into the subsequent fully connected layer for calculation. The subsequent fully connected layer outputs a four-dimensional net mass spectrometry response value vector after interference correction and a four-dimensional interference correction factor vector. The interference correction factor vector represents the correction multipliers calculated by the model to be applied to the original response values ​​of the four analytes. The concatenation method is to connect the two vectors end to end in sequence. The activation function of the subsequent fully connected layer can be the ReLU function.

[0035] This embodiment extracts features through dual paths and establishes the correlation between the features of the analyte and the interference by combining the cross-attention module. It can accurately capture the influence of the interference on different analytes. Compared with the existing technology, the feature extraction is more targeted and the interference correction is more accurate, which solves the problems of insufficient targeting and poor effect of model correction in the existing technology.

[0036] In another embodiment, the training method for the machine learning model includes: constructing a training dataset, which includes a synthetic spiked training set and a real clinical validation set; the synthetic spiked training set is obtained by adding different known concentrations of tacrolimus, cyclosporine, sirolimus, and everolimus standards to a blank matrix to form a set of basic calibration samples; further adding various interfering substance standards selected from an interfering substance database to the basic calibration samples, the types and concentrations of interfering substances being added being designed based on their clinically common ranges to form multiple sets of spiked interfering samples; the real clinical validation set is obtained by collecting clinical samples that have undergone quantitative detection using a validated reference method; defining a composite loss function L, the expression of which is: L = L MSE +λ*L REG； Among them, L MSE L is the mean square error between the interference-corrected net mass spectrum response vector output by the model and the corresponding training label; REG The L2 norm regularization penalty term is applied to the interference correction factor vector of the model's synchronous output, and λ is the regularization strength coefficient. During training, the machine learning model is first trained using a synthetic spiked training set, and the composite loss function L is minimized through an optimization algorithm. Then, the performance of the trained model is evaluated using a real clinical validation set.

[0037] For example, the training method for the machine learning model includes constructing a training dataset, which includes a synthetic spiked training set and a real clinical validation set. The purpose of the training dataset is to provide data support for model training, enabling the model to learn the rules of interference correction. The synthetic spiked training set is obtained by adding four analyte standards at different known concentrations to a blank matrix to form a set of basic calibration samples. The blank matrix can be a healthy human blood matrix without immunosuppressants, and the concentration of the analyte standards can be 2 ng / mL or 5 ng / mL. Multiple interference standards selected from an interference substance database are further added to the basic calibration samples. The types and concentrations of interference added are designed based on their clinically common ranges, forming multiple sets of spiked interference samples. The concentration of the interference added can be 1 ng / mL or 3 ng / mL. The real clinical validation set is obtained by collecting clinical samples that have undergone quantitative detection using a validated reference method, such as high-performance liquid chromatography or an officially recognized mass spectrometry method. λ is the regularization intensity coefficient, which can be 0.01 or 0.05. The composite loss function is used to evaluate the deviation between the model output and the true value, while preventing model overfitting. During training, the machine learning model is first trained using a synthetic spiked training set. The composite loss function L is minimized by an optimization algorithm, which can be either stochastic gradient descent or the Adam algorithm. Then, the trained model is evaluated using a real clinical validation set, and the evaluation metric can be the degree of agreement between the predicted and actual values.

[0038] This embodiment uses a training set that combines synthetic spikes with real clinical samples, along with a composite loss function that includes a regularization term. This allows the model to learn interference correction patterns and has good clinical applicability, while reducing the risk of overfitting. Compared with existing technologies, the model training is more comprehensive and the generalization ability is stronger, solving the problems of poor model applicability and easy overfitting in existing technologies.

[0039] In another embodiment, S5 includes: S51: Reading the interference correction factor vector output by the machine learning model in S4; if the values ​​of all four elements are within a preset reference range, proceed to S52; if the value of any element exceeds the reference range, proceed to S53; S52: For tacrolimus, cyclosporine, sirolimus, and everolimus, divide the net mass spectrometry response value obtained in S4 by the mass spectrometry response value of the corresponding internal standard to obtain the internal standard normalized response value; substitute the internal standard normalized response value of each analyte into the pre-established corresponding polynomial standard curve to calculate its concentration; S53: [The text abruptly ends here, likely due to an incomplete translation or source material.] For analytes whose values ​​in the interference correction factor vector exceed the baseline range, dynamic curve correction is performed. Specifically: S531: The pre-stored standard curve equation for the analyte is called, which is a polynomial standard curve; S532: Using the interference correction factor of the analyte output by the machine learning model as a dynamic parameter, the function shape of the standard curve equation is adjusted for adaptability through a predefined curve correction function to generate an instantaneous correction curve that is only used for the current sample; S533: The ratio of the net mass spectrometry response value of the analyte to the internal standard response value is substituted into the instantaneous correction curve to calculate its concentration.

[0040] For example, the steps for calculating the concentrations of four analytes in a blood sample based on the net mass spectrometry response value, the corresponding internal standard mass spectrometry response value, and a pre-established standard curve include: first, reading the interference correction factor vector output by the machine learning model. The interference correction factor vector is a four-dimensional vector, corresponding to the correction coefficients of the four analytes; determining whether the values ​​of the four elements are all within a preset reference range, which can be 0.8-1.2. This determination step is to determine whether dynamic curve correction is needed. If the values ​​of the four elements are all within the preset reference range, then for each of the four analytes, the obtained net mass spectrometry response value is divided by the corresponding internal standard mass spectrometry response value to obtain the internal standard normalized response value. The calculation process is: analyte net mass spectrometry response value / corresponding internal standard mass spectrometry response value; substituting the internal standard normalized response value of each analyte into the pre-established corresponding polynomial standard curve to calculate its concentration. If the value of any element exceeds the baseline range, dynamic curve correction is performed on the analyte corresponding to the element in the interference correction factor vector whose value exceeds the baseline range. Specifically, this involves calling the pre-stored standard curve equation for that analyte, which is a polynomial standard curve; using the interference correction factor of the analyte output by the machine learning model as a dynamic parameter; and adjusting the function form of the standard curve equation through a predefined curve correction function to generate an instantaneous correction curve that is only applicable to the current sample. The curve correction function can be f(x) = kx + b or f(x) = kx 2 +bx+c, where k is the interference correction factor; substitute the ratio of the net mass spectrometry response value of the analyte to the internal standard response value into the instantaneous correction curve to calculate its concentration.

[0041] In existing technologies, fixed standard curves are often used to calculate immunosuppressant concentrations, without considering the differences in interference levels among different samples. This leads to significant deviations in the concentration calculation results for some samples with strong interference. This embodiment selects either a fixed standard curve or a dynamic curve correction for concentration calculation based on whether the interference correction factor exceeds the baseline range. This allows for different calculation methods to be adapted to samples with varying degrees of interference. Compared to existing technologies, the concentration calculation is more targeted and accurate, solving the problem that fixed curve calculations in existing technologies are difficult to adapt to samples with different levels of interference.

[0042] In another embodiment, the type of interference is determined based on the compound category information of the effective potential interfering substances marked in S3. If the effective potential interfering substances are mainly endogenous substances, a first correction function is invoked to correct the polynomial standard curve of the analyte. The first correction function is configured to perform nonlinear correction for the low concentration segment of the curve. If the effective potential interfering substances are mainly exogenous drugs or metabolites, a second correction function is invoked to correct the polynomial standard curve. The second correction function is configured to perform nonlinear correction for the high concentration segment of the curve. The interference correction factor of the analyte output by the machine learning model is used as an input parameter and substituted into the invoked correction function to generate an instantaneous correction curve that is only used for the current sample.

[0043] For example, based on the compound category information of the effective potential interfering substances marked in the interfering substance database, the type of interference is determined. The compound category information refers to the substance category to which the interfering substance belongs recorded in the database. The purpose of this determination step is to determine which curve correction function to use for dynamic correction. If the effective potential interfering substances are mainly endogenous substances, which are substances produced by the human body itself, such as bilirubin, then the first correction function is called to correct the polynomial standard curve of the analyte. The first correction function is configured to perform nonlinear correction for the low concentration range of the curve. The first correction function can be f(x) = k(0.6x) 2 +0.3x)+b, where k is the interference correction factor. This function has a larger correction coefficient in the low-concentration range, which can accurately correct the interference effect of endogenous substances at low concentrations. If the effective potential interfering substances are mainly exogenous drugs or metabolites, and the exogenous drugs are drugs ingested in vitro, such as antibiotics, then the second correction function is called to correct the polynomial standard curve. The second correction function is configured to perform nonlinear correction for the high-concentration range of the curve. The second correction function can be f(x)=k(0.2x)+b. 2 +0.7x)+b, where k is the interference correction factor. This function has a larger correction coefficient in the high-concentration range, which can accurately correct the interference of exogenous substances at high concentrations. The interference correction factor of the analyte output by the machine learning model is used as the input parameter and substituted into the called correction function to generate an instantaneous correction curve applicable only to the current sample. The generation process involves substituting the interference correction factor into the correction function to determine the specific parameters of the function, thereby obtaining the instantaneous correction curve.

[0044] This embodiment selects the corresponding correction function according to the type of interference and optimizes the correction strategy for both intrinsic and extrinsic interference. It can achieve accurate correction for different types of interference. Compared with the existing technology, the correction is more targeted and has a stronger ability to adapt to different interference scenarios.

[0045] The following is a description of a specific embodiment.

[0046] I. Standard Curve The specific steps are as follows: Preparation of Standards: Pure tacrolimus, cyclosporine, sirolimus, and everolimus standards were selected, and a series of standard solutions of varying concentrations were prepared using a blank matrix (blood matrix from healthy individuals without immunosuppressants) as the dilution medium. Among them: The concentration gradients of tacrolimus, sirolimus, and everolimus were 2 ng / mL, 5 ng / mL, 10 ng / mL, and 20 ng / mL, respectively. The concentration gradients of cyclosporine A were 40 ng / ml, 100 ng / ml, 150 ng / ml, and 400 ng / ml.

[0047] Internal standard addition: Add an internal standard solution (containing ascomycin, cyclosporine-d4, sirolimus-d3, and everolimus-d4, all at a concentration of 8 ng / mL) to each concentration of the standard solution, ensuring that the internal standard concentrations of the standard and the sample are consistent. The volume of the internal standard added should be 20% of the volume of the standard solution.

[0048] After processing the above series of standard solutions according to the same pretreatment procedure as the samples, they were injected into a liquid chromatography-tandem mass spectrometry system and detected using the same chromatographic and mass spectrometric conditions as the samples. The analytical mass spectrometry response signal value of each concentration standard and the corresponding internal standard mass spectrometry response signal value were recorded.

[0049] For each concentration point, the detection data were normalized using internal standards, i.e., the ratio of the analytical mass spectrum response value to the corresponding internal standard mass spectrum response value (denoted as y) was calculated. The actual concentration of the standard was plotted on the x-axis (denoted as x), and the normalized internal standard response value y was plotted on the y-axis, using a quadratic polynomial (y=ax). 2 Linear regression is performed on the equation (a, b, and c are fitting coefficients), and the fitting parameters are optimized using the least squares method to obtain the polynomial standard curve equations and correlation coefficients for each of the four analytes. The R-squared value is required to be... 2 ≥0.995, to ensure that the curve fitting accuracy meets the detection requirements.

[0050] Prepare intermediate concentration quality control samples (8 ng / mL for tacrolimus, sirolimus, and everolimus, and 120 ng / mL for cyclosporine A), substitute them into the standard curve equation to calculate the concentration, verify that the relative deviation between the measured concentration and the theoretical concentration is ≤5%, and after confirming the reliability of the curve, store the curve equation for subsequent sample concentration calculation.

[0051] II. Experimental Group and Control Group (a) Experimental Group 1. Sample pretreatment: Take 200 μL of venous whole blood sample from organ transplant patients and mix it with 50 μL of internal standard solution containing ascomycin, cyclosporine-d4, sirolimus-d3, and everolimus-d4 (all at a concentration of 8 ng / mL). Add 600 μL of methanol (precipitant) and vortex mix at 2500 r / min for 1.5 minutes. Then centrifuge at 9000 r / min for 8 minutes and take the supernatant as the sample to be tested.

[0052] 2. Liquid Chromatography-Tandem Mass Spectrometry (LC-MS / MS): A liquid chromatography system with a 4.6 mm × 150 mm, 5 μm C18 reversed-phase column was used. Mobile phase A was an aqueous solution containing 0.1% formic acid and 2 mM ammonium acetate, and mobile phase B was a methanol solution containing 0.1% formic acid. The gradient elution program was 0–0.5 min, mobile phase A volume fraction 85%; 0.5–1.5 min, mobile phase A volume fraction 2%; 1.5–2.8 min, mobile phase A volume fraction 85%. The tandem mass spectrometry system used electrospray ionization (ESI) in positive ionization mode, with an ion source temperature of 380 °C, a spray voltage of 3800 V, and multiple reaction monitoring (MRM) mode for detection.

[0053] 3. Interference retrieval: A pre-constructed database of interfering substances is used, which includes endogenous substances (bilirubin, cholesterol), exogenous drugs (amoxicillin, fluconazole) and their metabolites (amoxicillin hydroxylation products, fluconazole demethylation products). The name, chemical formula, typical retention time range and characteristic parent ion-daughter ion pairs of each compound are recorded. The interference judgment threshold is set to 1.3 times the relative response intensity benchmark.

[0054] 4. Machine Learning Model: A supervised learning model is adopted. The input data includes the original quantitative ion pair response signals of four analytes, the corresponding internal standard quantitative ion pair response signals, the characteristic ion pair response signals of effective potential interfering substances, and the ratio of the response intensity of interfering substances to internal standards. The model contains a first path with two fully connected layers, a second path with two fully connected layers, and a cross-attention module. The attention weights are calculated using the Softmax function, and the activation function of subsequent fully connected layers is the ReLU function. The composite loss function has λ=0.03, and the training uses the stochastic gradient descent algorithm.

[0055] 5. Concentration Calculation: The interference correction factor baseline range is set to 0.8-1.2; for endogenous interference, the first correction function f(x)=k(0.6x) is called. 2 +0.3x)+b (k is the interference correction factor), when there is external interference, the second correction function f(x)=k(0.2x) is called. 2 +0.7x)+b, generate an instantaneous correction curve to calculate the concentration.

[0056] (II) Control Group Design Control group 1: No interfering substance database search was performed. The raw mass spectrometry response value of the sample to be tested was directly substituted into the standard curve to calculate the concentration. All other parameters (pretreatment, chromatographic and mass spectrometry conditions, standard curve type) were the same as those of the experimental group.

[0057] Control group 2: The interference retrieval method was the same as that of the experimental group. After identifying the effective interfering substances, the net response value was obtained by simply subtracting the original response value of the analyte from the response value of the interfering substance. The concentration was then calculated by substituting the net response value into the standard curve. The other parameters were the same as those of the experimental group.

[0058] III. Experimental Procedure Design 1. Sample Preparation: Thirty clinical samples were selected and divided into three groups of 10 samples each. The first group was supplemented with endogenous interfering agents (bilirubin 20 μmol / L, cholesterol 5 mmol / L), the second group was supplemented with exogenous interfering agents (amoxicillin 10 ng / mL, fluconazole 12 ng / mL), and the third group was supplemented with both endogenous and exogenous interfering agents. Blank spiked samples without additional interference were also prepared (tacrolimus fk506, cyclosporine A, sirolimus, everolimus; the concentration gradients of tacrolimus fk506, sirolimus, and everolimus were 2 ng / mL, 5 ng / mL, 10 ng / mL, and 20 ng / mL, and the concentration gradients of cyclosporine A were 40 ng / mL, 100 ng / mL, 150 ng / mL, and 400 ng / mL, 2 samples each) for method validation.

[0059] 2. Parallel testing: All prepared test samples were tested in parallel using the methods of experimental group, control group 1, and control group 2. Each group of samples was tested three times, and the concentration values ​​of the four immunosuppressants were recorded for each test.

[0060] 3. Data preprocessing: Remove outliers that deviate from the mean by more than 10% in each group of repeated tests (if the proportion of outliers exceeds 1 / 3, the sample is retested), and calculate the average value of the test results of each method for each group of samples as the final test value of the sample.

[0061] 4. Index Calculation: Using the results of the accredited high performance liquid chromatography reference method as the true value, calculate the relative deviation of the results of each method (|detected value - true value| / true value × 100%), which serves as the core index for evaluating the accuracy of detection and the effectiveness of correction.

[0062] IV. Comparison of data between the experimental group and the control group V. Conclusion Existing technologies generally suffer from two major flaws: First, conventional multi-component mass spectrometry detection schemes (control group 1) rely solely on chromatographic separation to reduce interference, lacking a dedicated interference retrieval and correction mechanism, making them unable to handle complex interferences such as co-elution, resulting in significant detection bias. Second, even when using a simple interference correction scheme (control group 2), correction is achieved only through a single subtraction logic of the original response value and the interference response value, without fully considering the complex relationships between interfering substances, analytes, and internal standards, resulting in limited correction effectiveness. Experimental data show that under various interference scenarios, the relative deviation of the experimental group was significantly lower than that of the two control groups: in endogenous interference scenarios, the average relative deviation of the experimental group was only 3.1%, while control groups 1 and 2 were as high as 12.4% and 7.5%, respectively; in exogenous interference scenarios, the average relative deviation of the experimental group was 3.0%, while control groups 1 and 2 were 13.1% and 8.1%, respectively; in mixed interference scenarios, the average relative deviation of the experimental group was 3.5%, while control groups 1 and 2 reached 15.2% and 9.7%, respectively. Furthermore, the deviation fluctuation range of the four substances in the experimental group was controlled within 0.7%, demonstrating significantly better stability than the control groups. This indicates that the technical approach adopted by the experimental group can achieve accurate cancellation of complex interference, effectively solving the core problems of insufficient interference identification, simple correction logic, and poor detection accuracy in existing technologies.

[0063] 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 method for the simultaneous and rapid detection of the concentrations of four immunosuppressants, characterized in that, include: S1: Mix the blood sample with an internal standard solution containing ascomycin, cyclosporine-d4, sirolimus-d3 and everolimus-d4, add a precipitant and centrifuge to obtain the sample to be tested; S2: Inject the sample to be tested into a liquid chromatography-tandem mass spectrometry system for analysis to obtain raw data on the retention time windows and mass spectrometry response signals of four analytes: tacrolimus, cyclosporine, sirolimus, and everolimus. S3: Retrieve the raw data of S2 based on a pre-built interference substance database. The interference substance database contains information on a variety of compounds and their characteristic ions that are known to cause mass spectrometry signal interference within the retention time windows of the four analytes. S4: Input the raw data obtained in S2 and the potential interfering information retrieved in S3 into a pre-trained machine learning model, and output the interference-corrected net mass spectrometry response values ​​for tacrolimus, cyclosporine, sirolimus and everolimus in the current sample. S5: Based on the net mass spectrometry response value obtained in S4, the corresponding internal standard mass spectrometry response value, and the pre-established standard curve, calculate the concentrations of tacrolimus, cyclosporine, sirolimus, and everolimus in the blood sample.

2. The method for simultaneous rapid detection of four immunosuppressant concentrations as described in claim 1, characterized in that, In S2, the sample to be tested is injected into a liquid chromatography system equipped with a C18 reversed-phase column for separation. Mobile phase A is an aqueous solution containing 0.1% formic acid and 2mM ammonium acetate, and mobile phase B is a methanol solution containing 0.1% formic acid. A gradient elution program is used. The effluent after chromatographic separation was fed into a tandem mass spectrometry system and detected using electrospray ionization mode and multiple reaction monitoring mode.

3. The method for simultaneous rapid detection of four immunosuppressant concentrations as described in claim 1, characterized in that, In S3, the interference substance database includes endogenous substances, exogenous drugs and their metabolites that have been observed to interfere with the mass spectrometry signal within the specific retention time windows of the four analytes; each compound is recorded with its name, chemical formula, typical retention time range, at least one pair of characteristic parent ion and daughter ion pairs and their relative response intensity reference information under the detection system; During the search, the ion pair signals monitored in the original data obtained by S2 within the retention time window of each analyte for the interference substance database are matched with the characteristic ion pairs of the interference substances stored in the interference substance database. For suspected interference substances that are successfully matched, the intensity of its mass spectrometry response signal in the current sample is further compared with the relative response intensity benchmark of the interference substance under the same conditions recorded in the database. If its response intensity exceeds the preset interference judgment threshold, the compound is marked as a valid potential interference substance.

4. The method for simultaneous rapid detection of four immunosuppressant concentrations as described in claim 3, characterized in that, In S4, the pre-trained machine learning model is a supervised learning model. The input data received by the input layer includes: (a) the original quantitative ion pair response signals obtained in S2 within their respective retention time windows for each of the four analytes: tacrolimus, cyclosporine, sirolimus, and everolimus; (b) the quantitative ion pair response signals obtained within the same retention time window for the internal standards corresponding to the four analytes; (c) the characteristic ion pair response signals monitored within the corresponding analyte retention time window for each compound marked as a potential interfering agent in S3; and (d) the ratio of the response intensity of each potential interfering agent to the corresponding analyte internal standard, calculated based on the signals in (b) and (c).

5. The method for simultaneously and rapidly detecting the concentrations of four immunosuppressants as described in claim 3, characterized in that, The machine learning model includes a first path, a second path, and a cross-attention module; The first path contains at least one fully connected layer, which is used to extract features of the standardized analyte response value and the internal standard response value to generate a first feature vector. The second path contains at least one fully connected layer, which is used to extract the characteristics of the normalized response values ​​of each effective potential interfering object and the ratio of their response intensity to the corresponding analyte internal standard, and generate a second feature vector. The cross-attention module uses the first feature vector generated by the first path as the query vector and the second feature vector generated by the second path as the key vector and value vector. By calculating the attention weight between the query vector and the key vector, the value vector is weighted and summed to generate a context-aware interference feature vector. The context-aware interference feature vector is concatenated with the first feature vector and input together into the subsequent fully connected layer for calculation. The subsequent fully connected layer outputs a four-dimensional net mass spectrometry response value vector after interference correction and a four-dimensional interference correction factor vector. The interference correction factor vector represents the correction multiplier calculated by the model and applied to the original response values ​​of the four analytes.

6. The method for simultaneously and rapidly detecting the concentrations of four immunosuppressants as described in claim 5, characterized in that, Training methods for machine learning models include: A training dataset was constructed, comprising a synthetic spiked training set and a real clinical validation set. The synthetic spiked training set was obtained by adding different known concentrations of tacrolimus, cyclosporine, sirolimus, and everolimus standards to a blank matrix to form a basic calibration sample. Further, various interfering substance standards selected from an interfering substance database were added to the basic calibration sample, with the types and concentrations of interfering substances designed based on their clinically common ranges, forming multiple sets of spiked interfering samples. The real clinical validation set was obtained by collecting clinical samples that had undergone quantitative detection using a validated reference method. Define the composite loss function L as: L = L MSE +λ*L REG； Among them, L MSE L is the mean square error between the interference-corrected net mass spectrum response vector output by the model and the corresponding training label; REG The L2 norm regularization penalty term is applied to the interference correction factor vector of the model's synchronous output, where λ is the regularization strength coefficient. During training, the machine learning model is first trained using a synthetic spiked training set, and the composite loss function L is minimized by an optimization algorithm. Then, the performance of the trained model is evaluated using a real clinical validation set.

7. The method for simultaneous rapid detection of four immunosuppressant concentrations as described in claim 5, characterized in that, S5 include: S51: Read the interference correction factor vector output by the machine learning model in S4. If the values ​​of all four elements are within the preset baseline range, proceed to S52; if the value of any element exceeds the baseline range, proceed to S53. S52: For tacrolimus, cyclosporine, sirolimus, and everolimus, the net mass spectrometry response value obtained in S4 is divided by the mass spectrometry response value of the corresponding internal standard to obtain the internal standard normalized response value; the internal standard normalized response value of each analyte is substituted into the pre-established corresponding polynomial standard curve to calculate its concentration. S53: For analytes whose median values ​​in the interference correction factor vector exceed the baseline range, perform dynamic curve correction; specifically: S531: Call the pre-stored standard curve equation of the analyte, which is a polynomial standard curve; S532: Using the interference correction factor of the analyte output by the machine learning model as a dynamic parameter, the function shape of the standard curve equation is adapted through a predefined curve correction function to generate an instantaneous correction curve that is only used for the current sample. S533: Substitute the ratio of the net mass spectrometry response value of the analyte to the internal standard response value into the instantaneous calibration curve to calculate its concentration.

8. The method for simultaneous rapid detection of four immunosuppressant concentrations as described in claim 7, characterized in that, Based on the compound category information of the effective potential interfering substances marked in S3, determine the type of interference; If the effective potential interfering substances are mainly endogenous substances, the first correction function is invoked to correct the polynomial standard curve of the analyte. The first correction function is configured to perform nonlinear correction for the low concentration range of the curve. If the effective potential interfering substances are mainly exogenous drugs or metabolites, the second correction function is invoked to correct the polynomial standard curve. The second correction function is configured to perform nonlinear correction for the high concentration segment of the curve. The interference correction factor of the analyte output by the machine learning model is used as the input parameter and substituted into the called correction function to generate an instantaneous correction curve that is only used for the current sample.

Images