A method for rapidly screening fentanyl compounds based on a machine learning model

Abstract

The present invention relates to a method for rapidly screening fentanyl compounds based on a machine learning model, comprising the following steps: 1) establishing an MS / MS mass spectrometry database of fentanyl compounds; 2) constructing a machine learning model for fentanyl compounds; 3) extracting fentanyl compounds from a sample to be tested and collecting MS / MS mass spectrometry data, and step 4) rapid screening and identification of fentanyl compounds. The machine learning model of the present invention can rapidly identify, screen and detect various fentanyl compounds (including known, unknown new substitutes and metabolites, etc.), providing effective technical support for the discovery and supervision of fentanyl compounds.

Description

Technical Field

[0001] The present invention belongs to the field of analytical detection, and particularly relates to a method for rapidly screening fentanyl compounds based on a machine learning model. Background Art

[0002] As an artificially synthesized opioid receptor agonist, fentanyl can selectively activate the μ receptor to produce an analgesic effect, and its analgesic effect is 50-100 times that of morphine. However, due to its similar effects to heroin, fentanyl derivatives have been continuously "designed" from underground laboratories and have become typical representatives of contraband. In China, fentanyl compounds are defined as substances that meet one or more of the following conditions compared with the chemical structure of fentanyl: (1) replacing the propionyl group with other acyl groups; (2) replacing the phenyl group directly connected to the nitrogen atom with any substituted or unsubstituted monocyclic aromatic group; (3) substituting the piperidine ring with alkyl, alkoxy, alkenyl, hydroxyl, ester, ether, halogen, haloalkyl, amino, nitro, etc.; (4) replacing the phenethyl group with any other group (except hydrogen atom). Based on its highly derivatized structural characteristics, even if existing drugs are listed as controlled substances, criminals can still modify their skeletons and continuously design and synthesize new fentanyl compounds not on the control list, which has led to an increasing number of deaths caused by the illegal use of fentanyl compounds and has caused great harm to society.

[0003] China has always attached great importance to the control of fentanyl substances. In the 1996 edition of the "Catalogue of Narcotic Drugs", 12 fentanyl substances were included in the national control list. Facing the continuous update of fentanyl substances, China has actively taken measures to control them. In 2017, four fentanyl substances, namely carfentanil, furanylfentanyl, acrylfentanyl, and valerylfentanyl, were included in the "Supplementary Catalogue of Controlled Non-Medicinal Narcotic and Psychotropic Substances". By the end of 2018, China had controlled 25 fentanyl substances and 2 fentanyl precursors. On May 1, 2019, China included fentanyl substances in the "Supplementary Catalogue of Controlled Non-Medicinal Narcotic and Psychotropic Substances", which means that not only a single fentanyl substance is controlled, but also substances with similar chemical structures and effects to fentanyl are controlled, so as to effectively address the problems of rapid mutation and difficult crackdown of fentanyl drugs.

[0004] For the detection of fentanyl compounds, targeted detection methods were used in the past. However, this detection method is mainly aimed at qualitative and quantitative detection of known fentanyl compounds, and it is still impossible to perform qualitative or quantitative detection on the emerging new fentanyl compounds. Therefore, the detection and supervision of these fentanyls have always been in a relatively backward state. In order to identify a wider range of fentanyl compounds and discover new fentanyl compounds that have not been reported before, researchers have gradually adopted non-targeted data collection methods, such as mass spectrometry detection technology, which has become the preferred non-targeted analysis and detection method for fentanyl compounds due to its fast analysis speed, high detection sensitivity and wide range of applications. However, many fentanyl compounds are not suitable for detection by gas chromatography-mass spectrometry, and the analysis effect is not good for thermally unstable or non-volatile compounds, so the detection effect and usage scenarios are limited. In addition, the data collected by non-targeted is massive and complex. How to find the required information in the complex data and how to parse the information are also the difficulties of non-targeted analysis at present.

[0005] In recent years, with the substantial increase in computer computing power, the field of machine learning has developed rapidly. Today, the application of machine learning models is no longer limited to classic problems such as computer vision, pattern recognition, and image segmentation, but has been widely infiltrated into data analysis in all walks of life. Compared with traditional data analysis methods, the biggest advantage of machine learning models is that they have the ability to automatically extract specific trends and features from massive data, fit the most appropriate function to the experimental data trend, and thus accurately predict new sample data. There are few studies on the use of machine learning methods for non-targeted rapid screening and identification of fentanyl compounds. Therefore, developing a method for detecting fentanyl compounds based on machine learning algorithms that can efficiently, accurately, and comprehensively detect all known and / or unknown fentanyl compounds, derivative compounds, and metabolites is a technical problem that needs to be solved in this field. Summary of the invention

[0006] In view of the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide a method for rapidly detecting fentanyl compounds based on a machine learning model, thereby providing a basis and support for the supervision of fentanyl compounds. The present invention uses liquid chromatography quadrupole time-of-flight tandem mass spectrometry (LC-QTOF MS) to establish an MS / MS mass spectrogram database of fentanyl compounds, uses a machine learning model to study the database, and constructs a Binning-RF two-classification model of fentanyl compounds; taking fentanyl compounds in urine as the main research object, the urine is pre-treated by auxiliary automatic magnetic solid phase extraction technology, and combined with the machine learning model of fentanyl compounds, the purpose of rapidly screening various known and unknown fentanyl compounds is achieved. The present invention can be beneficial to strengthening the supervision of illegal fentanyl compounds and improving the ability to prevent drug abuse.

[0007] To achieve the above object, the present invention provides a method for rapidly screening fentanyl compounds based on a machine learning model, comprising the following steps:

[0008] Step 1) Establish a MS / MS mass spectrometry database of fentanyl compounds

[0009] Collect the secondary mass spectrometry data of fentanyl compounds and perform data preprocessing to obtain a MS / MS mass spectrometry database of fentanyl compounds. The data sources of the MS / MS database of fentanyl compounds include two parts. One is the secondary mass spectrometry data collected in the laboratory, and the other is the valid secondary mass spectrometry data of fentanyl compounds in the network public database.

[0010] The secondary mass spectrometry data collected in the laboratory: Use a liquid chromatography quadrupole time-of-flight mass spectrometer (LC-QTOFMS mass spectrometer) to collect the MS / MS mass spectrometry data of known standard fentanyl compounds.

[0011] The chromatographic conditions are as follows: Mobile phase A is a combination of formic acid and ammonium formate aqueous solution, the volume ratio of formic acid to ammonium formate aqueous solution is (1-5):1000, and the concentration of ammonium formate in the ammonium formate aqueous solution is 10-15 mM; Mobile phase B is a combination of acetonitrile and mobile phase A, where the volume percentage of acetonitrile is 90-95%; Flow rate: 0.3-0.5 mL / min; Gradient elution is adopted, and the injection volume is controlled: 2-5 μL.

[0012] In an embodiment of the present invention, mobile phase A: 0.1% formic acid + 99.9 (v / v)% ammonium formate aqueous solution with a concentration of 10 mM, mobile phase B: a mixed solution of 95 (v / v)% acetonitrile + 5% mobile phase A (0.1% formic acid + 99.9 (v / v)% ammonium formate aqueous solution with a concentration of 10 mM).

[0013] In an embodiment of the present invention, the elution gradient: 5% B (0 min), 5% B (0.5 min), 15% A (1.5 min), 30% A (4 min), 30% A (7.5 min), 90% A (8 min), 90% A (10 min), 5% A (10.1 min), 5% A (12 min).

[0014] The mass spectrometry conditions are as follows: Ion source ESI+, curtain gas 30-35 psi, ionization voltage 5000-5500 kV, temperature 500-600 °C, nebulizing gas 50-55 psi, auxiliary heating gas 50-55 psi, declustering voltage 100-120 V, cone voltage 60-65 V, mass spectrometry acquisition mode is MS-IDA mode, the primary mass spectrometry acquisition range is 100-1000 m / z, and the secondary mass spectrometry acquisition range is 50-1000 m / z.

[0015] In one embodiment of the present invention, the ion source is ESI+, the curtain gas is 35 psi, the ionization voltage is 5500 kV, the temperature is 600 °C, the nebulizing gas is 55 psi, the auxiliary heating gas is 55 psi, the declustering voltage is 100 V, and the cone voltage is 65 V.

[0016] The MS / MS data of the network public database include the MS / MS mass spectrometry data of known fentanyl compounds in the National Institute of Standards and Technology (NIST), MassBank of North America (MoNA), and HighResNPS network databases.

[0017] The chemical formulas, InChIKeys, precursor ions, collision energies, molecular weights, etc. of the collected fentanyl compounds are entered into a self-built library. After data preprocessing steps such as normalization, data augmentation, and data cleaning, they are used as a positive data set, and an MS / MS mass spectrometry database of fentanyl compounds is constructed for machine learning.

[0018] Specifically, it includes the following processing steps:

[0019] 1-1) Data extraction: Use MS-DIAL 4.60 software to extract peaks from the wiff format file of the collected data. The parameter settings are as follows: MS1 tolerance 0.01 Da, MS2 tolerance 0.025 Da, Minimum peak height 1000 amplitude, Mass slice width 0.1 Da, Sigma window value 0.5, MS / MS abundance cut off 0 amplitude, Adduct ion included [M+H]+. After processing, export the spectral information.

[0020] 1-2) Normalization processing: Convert the absolute response intensity to the relative response intensity and normalize it to 1000 (the maximum relative response value) at the same time. The normalization formula is as (I), and perform smoothing or peak searching operations as needed.

[0021] Relative response value = (absolute response value / maximum absolute response value) × maximum relative response value (I)

[0022] Since the network public databases have different collection instruments and parameter settings, it is necessary to unify their formats. First, convert the absolute response intensity into relative response intensity according to the method described in the self-built database before, and then uniformly represent the mass spectrometry data set as L, L = {L1 / L2 / L3… / L N}, where each variable represents a mass spectrometry data, N is the total number of sample data, and each Lx consists of the fragment mass number m x and the relative response intensity i x . That is, Lx = (m x i x ), where m x = {m1, m2, m3…, m n}, i x = {i1, i2, i3…, i n}.

[0023] 1-3) Data augmentation: Adopt data augmentation means to reduce the problem of the performance degradation of the machine learning model caused by the imbalance between the positive data set and the negative data set.

[0024] For each original spectrogram, retain the spectrogram data with its maximum response intensity above x% (x = 50, 60, 70, 80, 90, 100) respectively for data augmentation to match the quantity of the subsequent negative data set.

[0025] 1-4) Data cleaning: Eliminate the spectrograms with less than 5 fragments after data augmentation to avoid the weak spectral characteristics affecting the final accuracy of the model, and complete the establishment of the MS / MS mass spectrometry database of fentanyl compounds.

[0026] Furthermore, collect blank urine, solvent blank, and high-quality spectrograms that are not classified as fentanyl in the Exposome-Explorer public database to jointly form a negative data set; use both the positive data set and the negative data set as the training set of the mass spectrometry database to train the machine learning model in step 3).

[0027] The solvent refers to all solvents used in the experimental process of the present invention, such as methanol and acetonitrile.

[0028] Step 2) Construct a machine learning model for fentanyl compounds

[0029] Through feature extraction, and use a random forest (RF) classification model to classify the data of fentanyl compounds.

[0030] 2-1) Feature extraction: Adopt the Top N method, Binning method or Grid method to extract and optimize the features of the MS / MS spectrograms.

[0031] Preferably, the Binning method is used to extract features from the MS / MS spectra and optimize the parameters.

[0032] The parameter optimization includes the optimization of the intensity threshold, the m / z range of fentanyl compounds, and the bin width.

[0033] 2-2) Model construction: Construct a random forest RF classification model for learning and classification.

[0034] Based on the CART (Classification and Regression Trees) algorithm, the present invention independently develops a random forest RF classification model. By constructing and evaluating decision trees, the selection of split points and pruning strategies are determined, thereby optimizing the structure and performance of the decision trees. The optimal features are selected with the help of the Gini index, and at the same time, the optimal binary split point of the feature is determined. The calculation formula of the Gini index is as described in (II):

[0035]

[0036] In the formula, Pi is the proportion of the i-th class samples in the dataset, and k is the total number of classes.

[0037] The prediction of the RF model is the voting result of all decision tree prediction results. Assuming that we have N decision trees, the probability of predicting the class as C can be expressed as formula (III):

[0038]

[0039] In the formula, I(C i = C) is an indicator function, indicating whether the prediction of the i-th decision tree is the class C.

[0040] Furthermore, the Binning-RF model is evaluated for its performance using 5*3 nested cross-validation and the hyperparameters of the model are optimized. Among them, the outer loop is used 5 times for algorithm evaluation, and the inner loop is used 3 times for Bayesian optimization of hyperparameter tuning.

[0041] Bayesian optimization is to, given an optimization objective function (a general function, only the input and output need to be specified, without knowing the internal structure and mathematical properties), update the posterior distribution (Gaussian process) of the objective function by continuously adding sample points until the posterior distribution basically fits the true distribution.

[0042] In an embodiment of the present invention, the objective function is the performance index of the classifier, such as accuracy, precision, recall rate, F1 score, Matthews correlation coefficient (MCC), ROC-AUC, etc. These indicators, as the output values of the objective function, can reflect the performance of the model on the test data.

[0043] The Bayesian optimization of hyperparameters includes optimizing the hyperparameters of n_estimators, max_features, max_depth, min_samples_split, min_samples_leaf, and bootstrap of the random forest RF model.

[0044] 2-3) Model evaluation: Use accuracy, precision, recall, F1 score, and MCC score (Matthew's correlation coefficient) as the evaluation metrics of the model.

[0045] Furthermore, considering that the training dataset of the present invention belongs to imbalanced data (i.e., negative data is greater than positive data), we calculate the Matthews correlation coefficient (MCC) score to examine the model prediction accuracy, and its calculation formula is as follows (IV):

[0046]

[0047] In the formula, TP is the true positive (TP): the positive sample predicted as positive by the model; FP is the false positive (FP): the negative sample predicted as positive by the model; FN is the false negative (FN): the positive sample predicted as negative by the model; TN is the true negative (TN): the negative sample predicted as negative by the model.

[0048] The present invention also compares the RF model with other machine learning algorithms such as LR, SVM, and KNN. The experimental results show that the RF classification model is significantly better than the support vector machine, K-nearest neighbor, and logistic regression. Therefore, it can be determined that the Binning-RF classification model of the present invention is effective and excellent.

[0049] In an embodiment of the present invention, the evaluation results of the Binning-RF classification model are: accuracy is 0.99±0.01, recall is 0.98±0.02, precision is 0.99±0.01, F1 is 0.99±0.01, and MCC is 0.98±0.01.

[0050] Step 3) Extract fentanyl-like compounds from the sample to be tested and collect MS / MS mass spectrometry data

[0051] 3-1) Extraction of fentanyl-like compounds:

[0052] Adopt the automatic magnetic extraction method to extract fentanyl compounds in the sample to be tested: sequentially activate the HLB magnetic solid-phase extraction material with pure methanol and 5v% methanol, and obtain it by treating with the eluent and the eluate.

[0053] In the activation step, add 2-3 mg of HLB magnetic solid-phase extraction material to every 400 uL of the sample;

[0054] The elution process is to perform elution treatment with two kinds of eluents respectively.

[0055] The composition of eluent 1 is: 5v% methanol aqueous solution, and the composition of eluent 2 is: 10%-20v% methanol aqueous solution;

[0056] The composition of the eluate is: 95v% methanol + 0.5v%-1v% formic acid aqueous solution;

[0057] The mixing time for each step is 30-60 s, and the magnetic adsorption time is 30-60 s.

[0058] After automatic magnetic extraction, re-dissolve with the initial mobile phase, centrifuge, and take the supernatant, which is the sample to be detected.

[0059] In an embodiment of the present invention, nitrogen blow-dry at 40 °C and centrifuge at 12000 rpm for 10 min are adopted.

[0060] In the prior art, the direct dilution method or the MCX solid-phase extraction method is commonly used for the extraction of fentanyl compounds. However, the above methods are all applicable to the extraction of known or target fentanyl compounds. In the case of facing numerous fentanyl derivatives or under the uncertain modification of the fentanyl molecular skeleton, using the above conventional extraction methods, it is impossible to comprehensively obtain the fragment fingerprint peaks of all fentanyl skeleton compounds, and the MS / MS spectrum is not accurate and comprehensive enough.

[0061] In the above conventional extraction methods, due to the certainty of the target compound, the mass spectrometry peaks of some unknown fentanyl compounds are easily destroyed or "sacrificed", and effective extraction cannot be achieved. Therefore, the above conventional extraction methods cannot identify all the fentanyl compounds in the sample to be tested, and cannot solve the supervision problems of various unknown fentanyl compounds. Moreover, even if the learning model of the present invention is adopted, it will not be quickly and effectively identified (see Comparative Examples 4-5).

[0062] The present invention adopts the HLB magnetic solid-phase extraction method. Through continuous optimization of this method, it can achieve the effective extraction of all fentanyl compounds in the sample to be tested, provide strong technical support for the identification and supervision of various fentanyl compounds, and on the basis of effective extraction, with the help of the machine learning model of the present invention, the specific structure and type of fentanyl derivatives can be quickly and accurately identified.

[0063] 3-2) Collect MS / MS mass spectrometry data of fentanyl-related compounds to be measured:

[0064] Use a liquid chromatography quadrupole time-of-flight mass spectrometer (LC-QTOF MS mass spectrometer) to collect MS / MS mass spectrometry data of fentanyl-related compounds.

[0065] Since it is impossible to enumerate all fentanyl-related compounds and their MS / MS mass spectrometry information directly affects the accuracy of the machine learning model, the selection of the mobile phase for the chromatographic conditions is crucial. The present invention innovatively uses a combination of specific mobile phases A and B to obtain more comprehensive and accurate MS / MS mass spectrometry information, which is crucial for the accuracy and effectiveness of subsequent machine learning.

[0066] The chromatographic conditions are as follows: Mobile phase A is a combination of formic acid and ammonium formate aqueous solution, the volume ratio of formic acid to ammonium formate aqueous solution is (1-5):1000, and the concentration of ammonium formate in the ammonium formate aqueous solution is 10-15 mM; Mobile phase B is a combination of acetonitrile and mobile phase A, where the volume percentage of acetonitrile is 90-95%; Flow rate: 0.3-0.5 mL / min; Gradient elution is used, and the injection volume is controlled at 2-5 μL.

[0067] The mass spectrometry conditions are as follows: Ion source ESI+, curtain gas 30-35 psi, ionization voltage 5000-5500 kV, temperature 500-600 °C, nebulizing gas 50-55 psi, auxiliary heating gas 50-55 psi, declustering voltage 100-120 V, cone voltage 60-65 V, mass spectrometry acquisition mode is MS-IDA mode, the primary mass spectrometry acquisition range is 100-1000 m / z, and the secondary mass spectrometry acquisition range is 50-1000 m / z.

[0068] In one embodiment of the present invention, the chromatographic conditions are as follows: Mobile phase A: 0.1 v% formic acid + 99.9 v% aqueous ammonium formate solution with a concentration of 10 mM; Mobile phase B: a mixed solution of 95 v% acetonitrile + 5 v% of Mobile phase A (0.1 v% formic acid + 99.9 v% aqueous ammonium formate solution with a concentration of 10 mM); Flow rate: 0.3 mL / min; Elution gradient: 5% B (0 min), 5% B (0.5 min), 15% A (1.5 min), 30% A (4 min), 30% A (7.5 min), 90% A (8 min), 90% A (10 min), 5% A (10.1 min), 5% A (12 min); Injection volume: 5 μL. The mass spectrometry conditions are as follows: Ion source ESI+, curtain gas 35 psi, ionization voltage 5500 kV, temperature 600 °C, nebulizing gas 55 psi, auxiliary heating gas 55 psi, declustering voltage 100 V, cone voltage 65 V, collision energy 30 V, mass spectrometry acquisition mode is MS-IDA mode, the first-stage mass spectrometry acquisition range is 100 - 1000 m / z, and the second-stage mass spectrometry acquisition range is 50 - 1000 m / z.

[0069] Further, the chromatographic conditions and mass spectrometry conditions are the same as those in step 1), and comprehensive mass spectrometry data of fentanyl compounds (including known and to-be-detected) can be obtained.

[0070] Step 4) Rapid screening and identification of fentanyl compounds

[0071] Use the machine learning model in step 2) to rapidly screen and classify the MS / MS mass spectrometry data of fentanyl compounds collected in step 3).

[0072] In one embodiment of the present invention, for butyrylfentanyl in urine matrix, the structure can be rapidly verified using the learning model of the present invention, indicating a high accuracy rate of the model.

[0073] In another embodiment of the present invention, an anesthetic fentanyl - fentanyl is incubated in vitro. After automatic magnetic extraction and treatment, its MS / MS mass spectrometry data is collected, and the machine learning model is used to identify and classify its in vitro incubation metabolites.

[0074] In the second aspect, the present invention provides a method for constructing a database of fentanyl compounds based on a machine learning model, including the following steps:

[0075] Step 1) Establish an MS / MS database of fentanyl compounds;

[0076] Step 2) Construct a machine learning model for fentanyl compounds and their metabolites.

[0077] The specific methods of steps 1) and 2) are the same as those in the first aspect.

[0078] Through the method of the present invention, a total of 772 MS / MS spectra of 277 fentanyl compounds, including 96 fentanyl compounds extracted in the laboratory, were collected and sorted out, and a secondary mass spectrometry database of fentanyl compounds was established using liquid chromatography-time-of-flight tandem mass spectrometry (LC-QTOFMS); the rapid automatic magnetic extraction pretreatment technology was used to effectively extract and purify both fentanyl compounds and their metabolites, and the spectral characteristics of fentanyl compounds were learned with the help of a machine learning model, so as to achieve the purpose of rapidly screening unknown fentanyl compounds.

[0079] Advantages of the present invention:

[0080] 1. For the first time, the present invention comprehensively collected and sorted out 772 MS / MS spectra of 277 fentanyl compounds, and established a secondary mass spectrometry database of fentanyl compounds using liquid chromatography-time-of-flight tandem mass spectrometry (LC-QTOF MS); this laid a comprehensive and powerful foundation for subsequent data identification and classification.

[0081] 2. The present invention uses a machine learning model to learn the spectral characteristics of fentanyl compounds, extracts features from MS / MS, and establishes an optimal Binning-RF classification model, so as to achieve the purpose of rapidly screening unknown fentanyl compounds.

[0082] 3. In view of the mass spectrometry data characteristics of fentanyl compounds, the present invention constructs a machine learning model, which is improved and customized on the basis of the RF model; this enables the model to have a high degree of adaptability and machine learning effect with fentanyl compounds, and improves the performance in specific application scenarios.

[0083] 4. The present invention uses the rapid automatic magnetic extraction pretreatment technology to effectively extract and purify the fentanyl compounds to be tested. Through effective pretreatment methods and optimization of liquid chromatography conditions, the solubility, separation effect and extraction efficiency of each fentanyl compound are improved. During the LC-QTOF MS detection process, the obtained secondary mass spectrometry (MS / MS) of each fentanyl compound has higher quality, more comprehensive compound types and better peak shapes. Furthermore, the training effect of the machine learning model is better, the prediction accuracy is higher, and the identification and screening of fentanyl compounds are realized efficiently, accurately and comprehensively.

[0084] 5. Using the machine learning model of the present invention, in addition to achieving efficient and accurate rapid screening of fentanyl compounds in common urine matrices, it can also accurately detect the metabolites of fentanyl, reflecting the accuracy and universality of the machine learning model of the present invention. It can be extended to the detection fields of any detection matrix, any fentanyl compound and its metabolites, providing guarantee for the control of fentanyl compounds. Description of the Drawings

[0085] Figure 1 This is the flowchart of the detection method in Embodiment 1 of the present invention.

[0086] Figure 2 This is the basic information of 24 fentanyl compounds in Embodiment 1.

[0087] Figure 3 This is the selection of the number of cut fragments in Embodiment 1.

[0088] Figure 4 This is the threshold selection in Embodiment 1.

[0089] Figure 5 This is the selection of the mass range and the number of bins in Embodiment 1.

[0090] Figure 6 This is the SHAP model explanation in Embodiment 1.

[0091] Figure 7 This is the result comparison of different feature extraction methods in Embodiments 1 - 3.

[0092] Figure 8 This is the extracted ion chromatogram (EIC) of butyrylfentanyl in urine samples in Embodiment 4.

[0093] Figure 9 This is the recognition result of butyrylfentanyl at different concentrations (1 ng / mL, 5 ng / mL) in urine by the machine learning model in Embodiment 4.

[0094] Figure 10 This is the recognition result of known fentanyl metabolites by the machine learning model in Embodiment 5.

[0095] Figure 11 This is the verification of new fentanyl metabolites in Embodiment 5; where a is the EIC diagram of the new metabolite standard; b is the EIC diagram of the fentanyl metabolite sample; c is the MS / MS spectrum of the new metabolite standard; d is the MS / MS spectrum of the fentanyl metabolite sample.

[0096] Figure 12 This is the extraction effect diagram of fentanyl compounds using MCX solid phase extraction material in Example 7 - A; where the left figure is the MS / MS spectrum of the extraction of fentanyl compounds at a low concentration point (0.5 ng / mL), and the right figure is the MS / MS spectrum of the extraction of fentanyl compounds at medium and high concentration points (10 ng / mL).

[0097] Figure 13 This is the extraction effect diagram of fentanyl compounds using HLB magnetic solid phase extraction material in Example 7 - B.

[0098] Figure 14 Extraction effect diagrams using different mobile phases for Example 8, where a is the MS / MS spectrum of Example 8-A; b is the MS / MS spectrum of Example 8-B; c is the MS / MS spectrum of Example 8-C.

[0099] Figure 15 Comparison diagram of the random forest RF classification model of Example 1 and other models of Comparative Examples 1-3. Detailed implementation manners

[0100] To make the objectives, technical solutions and advantages of the present invention clearer, the following further describes in detail the specific implementation manners of the present invention in conjunction with specific embodiments and the accompanying drawings.

[0101] The instrument and equipment used in the present invention are: ExionLC AD high performance liquid chromatograph - 7600 quadrupole time-of-flight mass spectrometer, AB SCIEX Company, USA.

[0102] The chemical reagents used in the experiments are all commercially available.

[0103] The urine sample matrix is from blank urine samples of healthy volunteers.

[0104] Example 1

[0105] A method for rapidly detecting fentanyl compounds based on a machine learning model. In this example, 24 fentanyl compounds are selected, an MS / MS mass spectrometry database is established, and a machine learning model for fentanyl compounds is constructed. The whole process is as Figure 1 shown.

[0106] 1) Establish an MS / MS database for fentanyl compounds

[0107] The secondary mass spectrometry data collected in the laboratory: The names, chemical formulas InChIKey, precursor ions, collision energies, MS / MS, adduct forms, retention times, etc. of Figure 2 the 24 fentanyl compounds are entered into a self-built library as reference standards; then the reference standard compounds in the self-built library are respectively prepared into standard stock solutions and working solutions, the chromatographic conditions and mass spectrometry conditions are adjusted, and the working solutions of the reference standards are used to collect secondary spectrum data in IDA mode.

[0108] Liquid chromatographic conditions and mass spectrometry conditions:

[0109] The chromatographic conditions were as follows: Mobile phase A: 0.1 v% formic acid + 99.9 v% aqueous solution of ammonium formate with a concentration of 10 mM; Mobile phase B: a mixed solution of 95 v% acetonitrile + 5 v% of Mobile phase A (0.1 v% formic acid + 99.9 v% aqueous solution of ammonium formate with a concentration of 10 mM); Flow rate: 0.3 mL / min; Elution gradient: 5% B (0 min), 5% B (0.5 min), 15% A (1.5 min), 30% A (4 min), 30% A (7.5 min), 90% A (8 min), 90% A (10 min), 5% A (10.1 min), 5% A (12 min); Injection volume: 5 μL.

[0110] The mass spectrometry conditions were as follows: Ion source ESI+, curtain gas 35 psi, ionization voltage 5500 kV, temperature 600 °C, nebulizer gas 55 psi, auxiliary heating gas 55 psi, declustering voltage 100 V, cone voltage 65 V, and fragmentation was carried out using collision energies of 10 V, 20 V, 30 V, and 40 V respectively. The mass spectrometry acquisition mode was MS-IDA mode, the first-order mass spectrometry acquisition range was 100 - 1000 m / z, and the second-order mass spectrometry acquisition range was 50 - 1000 m / z.

[0111] The mass spectrometry data of fentanyl-related compounds collected in the laboratory and the mass spectrometry data information in the network public database were input into a self-built library, and a MS / MS mass spectrometry database of fentanyl-related compounds was summarized to perform machine learning.

[0112] After organizing the data information in the database, it was used as the positive dataset for training. The organization included data extraction, normalization, data augmentation, data cleaning, and other data processing steps; blank urine, solvent blank, and high-quality spectra that were not classified as fentanyl in the Exposome-Explorer public database were collected to jointly form the negative dataset. Both the positive dataset and the negative dataset were used as the training sets to train the Binning-RF model in step 3).

[0113] 1-1) Data extraction: The obtained second-order mass spectrometry data was input into the self-built library. For the data collected in the laboratory, the MS-DIAL 4.60 software was used to extract peaks from the wiff format files collected, and the parameter settings were as follows: MS1 tolerance 0.01 Da, MS2 tolerance 0.025 Da, Minimum peak height 1000 amplitude, Mass slice width 0.1 Da, Sigma window value 0.5, MS / MS abundance cut off 0 amplitude, Adduct ion included [M+H]+. After processing, the spectral information was exported.

[0114] 1-2) Normalization: Among them, the MSMS spectrum column is two-dimensional information, including fragment mass numbers and absolute response intensities. For the convenience of subsequent machine learning steps, it is necessary to convert the absolute response intensity into a relative response intensity and normalize it to 1000 at the same time. The normalization formula is as follows:

[0115] Relative response value = (absolute response value / maximum absolute response value) × maximum relative response value

[0116] Among them, the maximum relative response value is 1000.

[0117] The dynamic background subtraction is selected in the data acquisition step, so these data do not need to be subjected to any smoothing or peak searching operations.

[0118] Since the network databases are different in terms of collection instruments and parameter settings, it is necessary to unify their formats. First, convert the absolute response intensity into a relative response intensity according to the method described in the self-built database before, and then uniformly represent the mass spectrometry data set as L, L = {L1 / L2 / L3… / L N}, where each variable represents a mass spectrometry data, N is the total number of sample data, and each Lx consists of fragment mass number m x and relative response intensity i x , that is, Lx = (m x i x ), where m x = {m1, m2, m3…, m n}, ix = {i1, i2, i3…, i n}.

[0119] 1-3) There are 772 spectra of 277 fentanyl compounds in the positive data set (96 were collected in the laboratory, and the rest are from network databases), but there are 4164 spectra in the negative data set. The number of the negative data set is much larger than that of the positive data set. This imbalance will cause the model to be biased towards the majority class during the training process, thus ignoring the characteristics of the minority class and resulting in a decline in the model performance. Therefore, to solve this data imbalance problem, we adopt data augmentation means to increase the size of the fentanyl spectral pool.

[0120] For each original spectrum, retain the data above x% (x = 50, 60, 70, 80, 90, 100) of its maximum response intensity respectively, and thus 3860 positive spectra are obtained.

[0121] 1-4) Data cleaning: For the positive and negative data sets, if the number of fragments in the spectrum is too small, it will lead to weak spectral characteristics and affect the final accuracy of the model. For example Figure 3As shown. Therefore, the number of fragments is checked, and spectra with less than 5 fragments are excluded; thus, the MS / MS mass spectrometry database of fentanyl compounds is established.

[0122] 2) Construction of the machine learning model for fentanyl compounds

[0123] Use the self-developed Random Forest (RF) classification model to perform learning classification on the positive and negative data sets in the fentanyl compound database:

[0124] 2-1) Feature extraction: Use the Binning method to extract features from the MS / MS mass spectrometry data and optimize the parameters of the Binning method. Since the mass spectrometry data contains a large amount of background noise and low-intensity signals, the noise may interfere with the true and effective signals. Therefore, it is necessary to filter out the noise below the threshold by setting the intensity threshold and retain the more meaningful high-intensity signals, thereby improving the quality of the data. At the same time, it can also reduce the amount of data to be processed, reduce the computational complexity and time, which is very important in machine learning training. Finally, because our data involves collection under different experimental conditions and different instruments and equipment, this may lead to large differences in signal intensity in the mass spectrometry data. By setting a unified intensity threshold, the data can be standardized and the uncertain differential features caused by external factors such as experimental conditions can be reduced.

[0125] Therefore, in this embodiment, the performance indicators of the model are investigated when the intensity thresholds are 10, 20, 30, and 40 respectively, and the results are as Figure 4 shown. It shows that when the intensity threshold is 10, the performance of the model is the best. The secondary mass spectrometry scanning range of fentanyl compounds in the self-built library is 50-1000 Da. Therefore, the data range is initially determined to be 50-1000 Da. However, after statistically analyzing the m / z range of the fragment ions of fentanyl compounds, it is found that 99.75% of the m / z ranges of fentanyl compounds are between 50-400 Da. Therefore, considering the computational amount, this embodiment also investigates the m / z range: 50-1000 and 50-400. The following two investigations are carried out on the bin width: 0.1 Da and 1 Da. The results of the four combinations of the m / z range and the bin width are as Figure 5 shown. Finally, the results of the training set show that the m / z range of 50-400 Da and the bin width of 0.1 Da are the optimal combinations.

[0126] 2-2) Model construction: Use the RF classification model to perform learning classification on the MS / MS mass spectrometry data after feature extraction.

[0127] By constructing and evaluating decision trees, the selection of splitting points and pruning strategies is determined to optimize the structure and performance of decision trees. The Gini index is used to select the optimal features and simultaneously determine the optimal binary splitting points for these features. The calculation formula of the Gini index is as described in (II):

[0128]

[0129] where Pi is the proportion of samples of the i-th class in the dataset and k is the total number of classes.

[0130] The prediction of the RF model is the voting result of all decision tree prediction results. Suppose we have N decision trees, and the probability of predicting class C can be expressed as formula (III):

[0131]

[0132] where I(C i = C) is an indicator function indicating whether the prediction of the i-th decision tree is class C.

[0133] Furthermore, 5 * 3 nested cross-validation is used to evaluate the performance of the model and optimize the model hyperparameters. Among them, the outer loop runs 5 times for algorithm evaluation, and the inner loop runs 3 times for Bayesian optimization of hyperparameter tuning, including the optimization of hyperparameters such as n_estimators, max_features, max_depth, min_samples_split, min_samples_leaf, and bootstrap for the random forest model.

[0134] 2-3) Model evaluation: The accuracy of the Binning-RF classification model is 0.99 ± 0.01, the recall rate is 0.98 ± 0.02, the precision is 0.99 ± 0.01, the F1 is 0.99 ± 0.01, and the MCC is 0.98 ± 0.01; finally, SHAP (SHapley Additive exPlanations) is used to interpret the prediction results of the Binning-RF model. SHAP is a model interpretation method that assigns importance values (Shap Values) to each feature for a specific prediction and identifies which feature is the most important, thereby facilitating the understanding of the decision-making process of machine learning models. It aims to provide transparency and interpretability for the predictions of machine learning models.

[0135] The results are as Figure 6The results show that the two fragment bins with the highest SHAP values are m / z 188.14 and m / z 105.07, and these two fragment ions happen to correspond to the characteristic fragments of fentanyl compounds. This not only explains why the Binning-RF model has a good classification effect on fentanyl compounds, but also proves from another perspective that this model has a good learning result on the fragmentation characteristics of fentanyl compounds.

[0136] Example 2

[0137] In step 2-1), the Top N method is used to replace the Binning method for feature extraction of MS / MS mass spectrometry data. The methods of other steps are the same as those in Example 1. Finally, the SHAP model is used to interpret the prediction results of the Top N-RF model.

[0138] Example 3

[0139] In step 2-1), the Grid-1D and Grid-2D methods are used to replace the Binning method for feature extraction of MS / MS mass spectrometry data. The methods of other steps are the same as those in Example 1. Finally, the SHAP model is used to interpret the prediction results of the Grid-RF model.

[0140] Figure 7 The results of using the Binning method, Top N method, Grid-1D and Grid-2D methods to extract features from MS / MS mass spectrometry data in Examples 1-3 are shown. By comparing the accuracy, precision, F1 score, and MCC score, it can be seen that the Binning method has higher accuracy, precision, F1 score, and MCC score, and is the optimal feature extraction method. In summary, the Binning-RF machine learning model is the best model for the identification and detection of fentanyl compounds.

[0141] Example 4 Screening and Identification of Butyrylfentanyl in Urine

[0142] The machine learning model constructed in Example 1 is used to quickly screen for fentanyl compounds - butyrylfentanyl in urine, providing a guarantee for the control of fentanyl compounds.

[0143] Among them, step 1-2) is the same as that in Example 1;

[0144] 3) Extract fentanyl compounds in the sample to be tested and collect MS / MS mass spectrometry data:

[0145] 3-1) Extraction of Fentanyl Compounds

[0146] Add butyrylfentanyl standard solution (100 μg / mL) to urine matrix to a final concentration of 5 ng / mL as the positive urine sample. Add 100 μL of HLB magnetic extraction material (20 mg / mL) to columns 1-6 of the automatic magnetic extraction device respectively, activation magnetic extraction solution: 100 μL of 5 v% methanol, 400 μL of urine sample; eluent 1: 400 μL of 5 v% methanol, eluent 2: 400 μL of 20 v% methanol, elution solution: 95 v% methanol + 0.5 v% formic acid. Mix for 60 s at each step, magnetic adsorption for 30 s, and automatically complete the magnetic extraction process. After magnetic extraction, transfer the eluate to another clean centrifuge tube, dry it with nitrogen at 40 °C, re-dissolve it with 100 μL of initial mobile phase, centrifuge at 12000 rpm for 10 min, and take the supernatant for injection to obtain the extract of fentanyl compounds in urine matrix.

[0147] 3-2) Collect MS / MS mass spectrometry data of fentanyl compounds

[0148] Use a liquid chromatography quadrupole time-of-flight mass spectrometer (LC-QTOF MS mass spectrometer) to collect MS / MS spectral data of fentanyl compounds.

[0149] Chromatographic conditions are as follows: mobile phase A: 0.1 v% formic acid + 99.9 v% aqueous ammonium formate solution with a concentration of 10 mM, mobile phase B: a mixed solution of 95 v% acetonitrile + 5 v% mobile phase A (0.1 v% formic acid + 99.9 v% aqueous ammonium formate solution with a concentration of 10 mM); flow rate: 0.3 mL / min; elution gradient: 5% B (0 min), 5% B (0.5 min), 15% A (1.5 min), 30% A (4 min), 30% A (7.5 min), 90% A (8 min), 90% A (10 min), 5% A (10.1 min), 5% A (12 min); injection volume: 5 μL.

[0150] Mass spectrometry conditions are as follows: ion source ESI+, curtain gas 35 psi, ionization voltage 5500 kV, temperature 600 °C, nebulizing gas 55 psi, auxiliary heating gas 55 psi, declustering voltage 100 V, cone voltage 65 V, collision energy 30 V, mass spectrometry acquisition mode is MS-IDA mode, primary mass spectrometry acquisition range 100 - 1000 m / z, secondary mass spectrometry acquisition range is 50 - 1000 m / z. The extracted ion chromatogram (EIC) of butyrylfentanyl in the urine sample is as Figure 8 shown.

[0151] 4) Screening and identification of fentanyl compounds

[0152] Perform non-targeted identification on the mixture of butyrylfentanyl standard and urine matrix through the machine learning model in step 2), and the identification results are as Figure 9As shown, it can be seen that the machine learning model can effectively identify butyrylfentanyl at different concentrations (1 ng / mL, 5 ng / mL) Figure 9 (in Figure 9 , Label being 1 indicates that it is a fentanyl compound), indicating that the model has high accuracy and precision.

[0153] Example 5 Screening and Identification of Fentanyl and Its in Vitro Incubated Metabolites

[0154] Using the machine learning model of Example 1, this example screens and detects fentanyl and its in vitro metabolites to provide assurance for the control of fentanyl compounds and their metabolites.

[0155] Among them, steps 1-2) are the same as in Example 1;

[0156] 3) Extract fentanyl compounds and their metabolites in the sample to be tested and collect MS / MS mass spectrometry data:

[0157] 3-1) Extraction of fentanyl compounds and their metabolites:

[0158] Use human liver microsomes (final concentration 0.5 mg protein / mL) to incubate the fentanyl standard solution (100 μg / mL, final concentration 5 μM) in vitro. The incubation system is 200 μL, which includes uridine 5′-diphosphate-glucuronic acid (UDPGA) (final concentration 5 mM), nicotinamide adenine dinucleotide phosphate (NADPH) (final concentration 5 mM), and reduced glutathione (GSH) (final concentration 10 μM). At 60 min of incubation, add 600 μL of ice-cold acetonitrile to terminate the reaction, vortex for 1 min, centrifuge at 16000 g at low temperature for 20 min, take the supernatant and evaporate it under a nitrogen stream at 30 °C. After drying, re-dissolve it with 400 uL of water to obtain the incubated sample.

[0159] Add 100 μL of HLB magnetic extraction material (20 mg / mL) to columns 1-6 of the automatic magnetic extraction device respectively, activate the magnetic extraction solution: 100 μL of 5% methanol, 400 μL of the incubated sample; eluent 1: 400 μL of 5 v% methanol, eluent 2: 400 μL of 20 v% methanol, and elution solution: 95 v% methanol + 0.5 v% formic acid. Mix for 60 s at each step, perform magnetic adsorption for 30 s, and automatically complete the magnetic extraction process. After magnetic extraction, transfer the eluate to another clean centrifuge tube, dry it under a nitrogen stream at 40 °C, re-dissolve it with 100 μL of the initial mobile phase, centrifuge at 12000 rpm for 10 min, and take the supernatant for injection to obtain the extract of fentanyl compounds and their metabolites in the incubated sample.

[0160] 3-2) Collection of MS / MS mass spectrometry data of fentanyl compounds and their metabolites

[0161] Using a quadrupole time-of-flight mass spectrometer (LC-QTOF MS mass spectrometer), under chromatographic and mass spectrometry conditions, different collision energies were used in the Information Dependent Analysis (IDA) mode to collect MS / MS spectral data of fentanyl compounds and their metabolites.

[0162] Chromatographic conditions are as follows: Mobile phase A: 0.1 v% formic acid + 99.9 v% aqueous ammonium formate solution with a concentration of 10 mM; Mobile phase B: a mixed solution of 95 v% acetonitrile + 5 v% of Mobile phase A (0.1 v% formic acid + 99.9 v% aqueous ammonium formate solution with a concentration of 10 mM); Flow rate: 0.3 mL / min; Elution gradient: 5% B (0 min), 5% B (0.5 min), 15% A (1.5 min), 30% A (4 min), 30% A (7.5 min), 90% A (8 min), 90% A (10 min), 5% A (10.1 min), 5% A (12 min); Injection volume: 5 μL.

[0163] Mass spectrometry conditions are as follows: Ion source ESI+, curtain gas 35 psi, ionization voltage 5500 kV, temperature 600 °C, nebulizing gas 55 psi, auxiliary heating gas 55 psi, declustering voltage 100 V, cone voltage 65 V, collision energy 30 V, mass spectrometry acquisition mode is MS-IDA mode, primary mass spectrometry acquisition range 100 - 1000 m / z, secondary mass spectrometry acquisition range is 50 - 1000 m / z.

[0164] 4) Application of the screening and identification model for fentanyl compounds and their metabolites

[0165] The machine learning model in step 2) was used for non-targeted identification of fentanyl compounds and in vitro incubated metabolites. The identification results are as Figure 10 shown. It can be seen that the model successfully identified and classified the known metabolites of fentanyl as fentanyl compounds, and also identified a new fentanyl metabolite that has not been reported in the literature (i.e., the compound labeled as "Unknown" in Figure 10 ).

[0166] To further verify the metabolites of this new fentanyl compound, we obtained the corresponding reference standards for compound verification. The results are as Figure 11 shown. Among them, a is the EIC map of the new metabolite reference standard; b is the EIC map of the fentanyl metabolite sample. It is speculated that the metabolite with a retention time RT of 4.503 is the metabolite of the new fentanyl compound, and its chemical structure is speculated.

[0167] Compound synthesis was carried out according to the speculated metabolite chemical structure. As metabolite standards, the metabolite standards and the novel metabolites in the incubation samples were respectively detected by the method in step 3) to obtain MS / MS spectra (Figures c and d). By comparing the metabolite standards in Figure c with the MS / MS spectrum of fentanyl in vitro incubation metabolites in Figure d, it can be found that the mass spectra of the two are basically consistent, indicating that the novel metabolites successfully predicted by the machine learning model of the present invention are determined to be a new type of fentanyl compound, and the model is also very accurate in predicting unknown fentanyl compounds.

[0168] The present invention first identified novel fentanyl compounds, expanded the exploration and identification of unknown fentanyl compounds, enriched the identification scope of fentanyl compounds, provided a strong technical basis for the control of such compounds, and facilitated the popularization and use in related fields.

[0169] Example 7 investigated the influence of different methods on the extraction of fentanyl compounds

[0170] Example 7-A MCX solid-phase extraction method

[0171] 1) Twenty-four fentanyl compounds in Example 1 were added to the urine matrix to obtain a solution with a final concentration of 100 ng / ml, and extraction was carried out using MCX solid-phase extraction material: the MCX magnetic solid-phase extraction material was activated successively with pure methanol and 5v% methanol, and treated with eluent and eluate, and re-dissolved with the initial mobile phase (95% mobile phase A + 5% mobile phase B, the composition of mobile phases A and B is the same as that in Example 1), centrifuged, and the supernatant was taken as the sample to be detected.

[0172] Among them, the eluent, eluate, and the mixing and treatment time were the same as those in step 3-1) of Example 4.

[0173] 2) Collect MS / MS mass spectrometry data:

[0174] Use a liquid chromatography quadrupole time-of-flight mass spectrometer (LC-QTOF MS spectrometer) to collect the MS / MS mass spectrometry data of fentanyl compounds.

[0175] The chromatographic conditions were as follows: Mobile phase A: 0.1 v% formic acid + 99.9 v% aqueous ammonium formate solution with a concentration of 10 mM; Mobile phase B: a mixed solution of 95 v% acetonitrile + 5 v% of Mobile phase A (0.1 v% formic acid + 99.9 v% aqueous ammonium formate solution with a concentration of 10 mM); Flow rate: 0.3 mL / min; Elution gradient: 5% B (0 min), 5% B (0.5 min), 15% A (1.5 min), 30% A (4 min), 30% A (7.5 min), 90% A (8 min), 90% A (10 min), 5% A (10.1 min), 5% A (12 min); Injection volume: 5 μL.

[0176] The mass spectrometry conditions were as follows: Ion source ESI+, curtain gas 35 psi, ionization voltage 5500 kV, temperature 600 °C, nebulizing gas 55 psi, auxiliary heating gas 55 psi, declustering voltage 100 V, cone voltage 65 V, collision energy 30 V, mass spectrometry acquisition mode was MS-IDA mode, the first-stage mass spectrometry acquisition range was 100 - 1000 m / z, and the second-stage mass spectrometry acquisition range was 50 - 1000 m / z.

[0177] The chromatogram data acquisition results of Example 7-A are shown in Figure 12 , and it can be seen that in the left figure, when using the MCX solid-phase extraction cartridge, for some low-concentration points (0.5 ng / mL) of fentanyl compounds (such as valerylfentanyl, α-methylthiofentanyl), effective extraction could not be carried out; in the right figure, when using the MCX solid-phase extraction cartridge, there was a tailing effect at some medium and high-concentration points (10 ng / mL). Therefore, the MCX solid-phase extraction material was not suitable for the comprehensive and effective extraction of fentanyl compounds, which was not conducive to the data acquisition and model training and learning of the subsequent machine learning model, and reduced the effectiveness of fentanyl compound identification.

[0178] Example 7-B HLB magnetic solid-phase extraction method

[0179] 1) Add the 24 fentanyl compounds in Example 1 to the urine matrix to obtain a solution with a final concentration of 100 ng / ml, and extract it using the HLB magnetic solid-phase extraction material. Other steps were the same as those in Example 7-A. The extraction effect diagram of fentanyl compounds is as shown in Figure 13 , and it can be seen that the MS / MS chromatogram peak shapes of the 24 fentanyl compounds were more complete, of high quality, and without tailing phenomenon, indicating that the HLB magnetic solid-phase extraction method was more suitable for the extraction of fentanyl compounds. Under the condition of complete and accurate peak shapes, it was more conducive to the training and learning of the subsequent machine model, so as to obtain an accurate and efficient learning model for the identification and screening of various fentanyl compounds.

[0180] Example 7-C Direct Dilution Method

[0181] The 24 fentanyl compounds in Example 1 were extracted using the direct dilution method. Specifically, the corresponding volumes of the 24 fentanyl standards were added to 1 mL of urine, and the samples were serially diluted to concentrations of 10 ng / mL, 5 ng / mL, 4 ng / mL, 3 ng / mL, 2 ng / mL, 1 ng / mL, and 0.5 ng / mL. Then, 2 mL of water was added for dilution, and the mixture was homogenized and dried with nitrogen. Next, 250 μL of the initial mobile phase (95% mobile phase A + 5% mobile phase B, where the compositions of mobile phases A and B are the same as those in Example 1) was added, and after homogenization, the mixture was centrifuged at 12,000 rpm for 10 min, and the supernatant was taken for injection analysis. The test conditions of the liquid chromatography quadrupole time-of-flight mass spectrometer (LC-QTOFMS) were the same as those in Example 7-A.

[0182] The results showed that some fentanyl compounds in urine, such as acetylfentanyl, ocfentanil, remifentanil, and cyclopropylfentanyl, could not be effectively extracted. In addition, Table 1 also shows the limits of detection (LOD) of various fentanyl compounds that could be accurately identified by the Binning-RF machine learning model in Example 1 after using the dilution method of Example 7-C and the HLB magnetic solid-phase extraction method of Example 1, respectively. The results indicated that using the magnetic solid-phase extraction method of Example 1 could significantly reduce the LOD of the machine learning model, had better compatibility with the machine learning model Binning-RF of the present invention, and was fully capable of detecting trace or micro amounts of fentanyl compounds; while the dilution method of Example 7-C had greater limitations and could not guarantee the accuracy and effectiveness of detecting micro amounts of fentanyl compounds.

[0183] Table 1

[0184]

[0185]

[0186] Example 8 Investigating the Extraction Effect of Different Chromatographic Mobile Phases on Fentanyl Compounds

[0187] Example 8-A

[0188] 1) Take the methanol standard solutions of 24 fentanyls, mix them to make the concentration 100 ng / mL;

[0189] 2) The chromatographic conditions are as follows:

[0190] Mobile phase A: Mobile phase A is an aqueous solution containing 0.1 v% formic acid, and mobile phase B is methanol; Flow rate: 0.3 mL / min; Elution gradient: 5% B (0 min), 5% B (0.5 min), 15% A (1.5 min), 30% A (4 min), 30% A (7.5 min), 90% A (8 min), 90% A (10 min), 5% A (10.1 min), 5% A (12 min); Injection volume: 5 μL.

[0191] The mass spectrometry conditions are as follows:

[0192] Ion source ESI+, curtain gas 35 psi, ionization voltage 5500 kV, temperature 600 °C, nebulizing gas 55 psi, auxiliary heating gas 55 psi, declustering voltage 100 V, cone voltage 65 V, fragmentation is carried out using collision energies of 10 V, 20 V, 30 V, and 40 V respectively. The mass spectrometry acquisition mode is MS-IDA mode, the first-stage mass spectrometry acquisition range is 100 - 1000 m / z, and the second-stage mass spectrometry acquisition range is 50 - 1000 m / z. The extraction effect diagrams of each fentanyl compound are shown in Figure 14 a.

[0193] Example 8-B

[0194] Only change the composition of mobile phases A and B in the chromatographic conditions in step 2), and other steps are the same as those in Example 8-A.

[0195] Among them, mobile phase A is 0.1 v% formic acid + 99.9 v% aqueous solution of ammonium formate with a concentration of 10 mM, and mobile phase B: acetonitrile. The extraction effect diagrams of each fentanyl compound are shown in Figure 14 b.

[0196] Example 8-C

[0197] Only change the composition of mobile phases A and B in the chromatographic conditions in step 2), and other steps are the same as those in Example 8-A.

[0198] Among them, mobile phase A is 0.1 v% formic acid + 99.9 v% aqueous solution of ammonium formate with a concentration of 10 mM, and mobile phase B: a mixed solution of 95 v% acetonitrile + 5 v% mobile phase A (0.1 v% formic acid + 99.9 v% aqueous solution of ammonium formate with a concentration of 10 mM). The extraction effect diagrams of each fentanyl compound are shown in Figure 14 c.

[0199] According to Figure 14It can be seen that, compared with Example 8-A and Example 8-B, Example 8-C provides a more comprehensive and higher-quality extraction of various fentanyl compounds. For example, the peak shapes of fentanyl, sufentanyl, 4-fluoroisobutyrlfentanyl, 3-methylthiofentanyl, and butyrfentanyl are better. Correspondingly, the quality of the secondary mass spectrometry (MS / MS) spectra of each fentanyl compound obtained is higher, which further improves the training effect of the machine learning model and the prediction accuracy. Therefore, it can be determined that the specific combination of mobile phases used in the present invention can obtain more comprehensive and high-quality MS / MS mass spectrometry information, whether in the construction of the self-built library of fentanyl compounds or in the detection process of the compound to be tested. By loading an accurate machine learning model, it can optimize the identification and screening of various fentanyl compounds.

[0200] Comparative Example 1

[0201] In step 2-2), the RF classification model in Example 1 was replaced with an LR classification model, and the other steps were the same as in Example 1.

[0202] Comparative Example 2

[0203] In step 2-2), the RF classification model in Example 1 was replaced with an SVM classification model, and the other steps were the same as in Example 1.

[0204] Comparative Example 3

[0205] In step 2-2), the RF classification model in Example 1 was replaced with a KNN classification model, and the other steps were the same as in Example 1.

[0206] Figure 15 Shows a comparison of the results of the machine learning models LR, SVM, and KNN in Comparative Examples 1-3 with the random forest model RF in Example 1. By comparing the results such as support vector machine, K-nearest neighbor, logistic regression, F1 score, and MCC score, it was determined that the results of the random forest model were the best. Combining with the method of feature extraction, the optimal classification model Binning-RF was established.

[0207] It can be seen from the above-mentioned examples and comparative examples that the present invention can effectively extract various fentanyl compounds (including known, unknown derivatives, metabolites, and compounds with skeleton modifications, etc.) by using a specific extraction process of fentanyl compounds and liquid chromatography detection conditions, and cooperate with a specific Binning-RF model to perform machine learning and classification on MS / MS mass spectrometry data. Through methods such as nested cross-validation and Bayesian optimization, while evaluating the accuracy of the model, the hyperparameters are optimized, and an efficient and accurate machine learning model is constructed for the first time. By comparing with the model, the present invention can quickly identify, screen, and detect various fentanyl compounds (including known, unknown derivatives, metabolites, and compounds with skeleton modifications, etc.); it provides effective technical support for the discovery, identification, and supervision of fentanyl compounds.

[0208] The above content is only the preferred embodiment of the present invention and is not used to limit the implementation of the present invention. Those of ordinary skill in the art can easily make corresponding modifications or alterations according to the main idea and spirit of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope required by the claims.

Claims

1. A method for rapid screening of fentanyl compounds based on a machine learning model, characterized in that: The steps include: 1) Collect the secondary mass spectrometry data of fentanyl compounds and perform data preprocessing to obtain the MS / MS mass spectrometry database of fentanyl compounds: 1-1) Data extraction; 1-2) Normalization: MS / MS mass spectrum data contains fragment mass number and absolute response intensity. The absolute response intensity is converted into relative response intensity and normalized to the maximum relative response value. The normalization formula is as follows (I): Relative response value = (absolute response value / maximum absolute response value) × maximum relative response value (I); 1-3) Data amplification: For each original spectrum of the positive data set, retain the spectrum data of x% of its maximum response intensity, perform data amplification, and match the number of negative data sets; where x = 50, 60, 70, 80, 90 or 100; 1-4) Data cleaning: After data amplification, the spectra with fragment numbers less than 5 were removed, and the MS / MS mass spectrum database of fentanyl compounds was established; Among them, a liquid chromatography quadrupole time-of-flight mass spectrometer was used to collect MS / MS mass spectrometric data of known standard fentanyl compounds; The chromatographic conditions are as follows: mobile phase A is a combination of formic acid and an aqueous solution of ammonium formate, wherein the volume ratio of formic acid to the aqueous solution of ammonium formate is (1-5):1000, and the concentration of ammonium formate in the aqueous solution of ammonium formate is 10-15 mM; mobile phase B is a combination of acetonitrile and mobile phase A, wherein the volume percentage of acetonitrile is 90-95%; 2) Building a machine learning model for fentanyl compounds: 2-1) Feature extraction: The Binning method is used to extract features and optimize parameters of MS / MS mass spectrometry data; the parameters include intensity threshold, m / z range of fentanyl compound precursor ions, and bin width; 2-2) Model construction: Based on the CART algorithm, the RF classification model was developed. The split point selection and pruning strategy were determined by building and evaluating the decision tree. The optimal feature was selected with the help of the Gini index, and the optimal binary split point of the feature was determined. The performance of the model was evaluated using 5*3 nested cross validation, and the Bayesian hyperparameters of the model were optimized. 2-3) Model evaluation: Use accuracy, precision, recall, F1 score and MCC score as evaluation indicators of the model; 3) Extract fentanyl compounds from the sample to be tested and collect MS / MS mass spectrometry data: 3-1) The fentanyl compounds in the sample to be tested are extracted by automatic magnetic extraction method; the magnetic solid phase extraction material is HLB; 3-2) Use a liquid chromatography quadrupole time-of-flight mass spectrometer to collect MS / MS mass spectrometric data of fentanyl compounds; 4) Rapid screening and identification of fentanyl compounds.

2. The method according to claim 1, characterized in that Step 1) the secondary mass spectrometry data of fentanyl compounds, including secondary mass spectrometry data collected in the laboratory and valid secondary mass spectrometry data on fentanyl compounds in the public database on the Internet; Secondary mass spectrometry data collected in the laboratory: Using a liquid chromatography quadrupole time-of-flight mass spectrometer, MS / MS mass spectrometry data of known fentanyl compounds were collected; Secondary mass spectrometry data from online public databases: including MS / MS mass spectrometry data of known fentanyl compounds in the National Institute of Standards and Technology, MassBank of North America, and HighResNPS online databases.

3. The method according to claim 2, characterized in that The mass spectrometry conditions were as follows: ion source ESI+, curtain gas 30-35psi, ionization voltage 5000-5500 kV, temperature 500-600°C, spray gas 50-55psi, auxiliary heating gas 50-55psi, declustering voltage 100-120 V, cone voltage 60-65 V; the mass spectrometry acquisition mode was MS-IDA mode, the primary mass spectrometry acquisition range was 100-1000 m / z, and the secondary mass spectrometry acquisition range was 50-1000 m / z.

4. The method according to claim 1, characterized in that The chemical formula, InChIKey, precursor ion, collision energy, molecular weight, MS / MS mass spectrum data, adduct form and retention time of the collected fentanyl compounds were entered into the self-built library. After normalization, data amplification and data cleaning preprocessing steps, an MS / MS mass spectrum database of fentanyl compounds was formed.

5. The method according to claim 1, characterized in that 1-1) Data extraction: Use MS-DIAL 4.60 software to perform peak extraction on the acquired wiff format files.

6. The method according to claim 1, characterized in that 2-2) The calculation formula of the Gini index is as described in (II): (II) Where Pi is the proportion of samples of the i-th category in the data set, and k is the total number of categories; The prediction of the RF model is the voting result of the prediction results of all decision trees. When there are N decision trees, the probability of predicting category C is expressed as formula (III): (III) In the formula, Is an indicator function, indicating whether the prediction of the i-th decision tree is category C.

7. The method according to claim 1, characterized in that The outer loop is used for 5 times to evaluate the model algorithm, and the inner loop is used for 3 times to optimize the hyperparameters in Bayesian way. The hyperparameters include: n_estimators, max_features, max_depth, min_samples_split, min_samples_leaf and bootstrap of the random forest RF model.

8. The method according to claim 1, characterized in that Step 3-1) using pure methanol and 5% methanol to activate the magnetic solid phase extraction material in sequence, and using eluent and eluting solution to obtain an extract of fentanyl compounds; 3-2) Chromatographic conditions are as follows: mobile phase A is a combination of formic acid and an aqueous solution of ammonium formate, wherein the volume ratio of formic acid to the aqueous solution of ammonium formate is (1-5):1000, and the concentration of ammonium formate in the aqueous solution of ammonium formate is 10-15 mM; mobile phase B is a combination of acetonitrile and mobile phase A, wherein the volume percentage of acetonitrile is 90-95%; flow rate: 0.3-0.5 mL / min; gradient elution is used, and the injection volume is controlled to be 2-5 μL; The mass spectrometry conditions were as follows: ion source ESI+, curtain gas 30-35psi, ionization voltage 5000-5500 kV, temperature 500-600°C, spray gas 50-55psi, auxiliary heating gas 50-55psi, declustering voltage 100-120 V, cone voltage 60-65 V; the mass spectrometry acquisition mode was MS-IDA mode, the primary mass spectrometry acquisition range was 100-1000 m / z, and the secondary mass spectrometry acquisition range was 50-1000 m / z.

9. Use of the method according to any one of claims 1 to 8 in the detection and identification of fentanyl compounds.

Images