Untargeted identification method and system for unknown pollutants based on mass spectrum and generative model

Abstract

The invention discloses an unknown pollutant non-target identification method and system based on mass spectrum and a generative model, and is applied to the technical field of environmental monitoring and analytical chemistry. The method comprises the following steps: collecting mass spectrum data of a water body sample to be detected; preprocessing the original mass spectrum data, and outputting standardized features; inputting the standardized features into a pre-trained generative model to generate a plurality of candidate molecular formulas; using chemical and physical constraints to eliminate candidate molecular formulas which do not conform to rules; executing a rule-driven algorithm to obtain candidate pollutant molecular structures; carrying out comprehensive scoring and sorting on candidate pollutant molecular structures through chemical prior and environmental prior; and semi-quantitative or relatively quantitative concentration determination is carried out. According to the method, a rule-driven expert system and a data-driven generation model are combined, unknown pollutants which do not exist in a standard library can be effectively recognized and analyzed, and full-process automatic processing from original mass spectrum data to pollutant structures and concentrations is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of environmental monitoring and analytical chemistry, and more particularly to a method and system for unknown pollutant non-target identification based on mass spectrometry and generative model. BACKGROUND

[0002] Organic pollutants in water bodies are extremely complex in species, including industrial emissions, agricultural non-point source pollution, domestic sewage, and drug residues, etc. They are characterized by a large number of species and rapid updates. In addition to common polycyclic aromatic hydrocarbons, pesticides, and surfactants, they also include emerging pollutants such as drugs and metabolites, personal care product ingredients, and new synthetic chemicals. The concentration range of these substances in the water environment spans from ng / L trace level to mg / L high concentration, and is significantly disturbed by complex matrices such as dissolved organic matter and salt. Existing detection methods mainly rely on targeted detection, i.e. establishing a calibration curve through standard samples to achieve accurate qualitative and quantitative analysis. However, this method requires the establishment of detection conditions for each pollutant, which is time-consuming and labor-intensive, and it is difficult to cover unknown pollutants or newly emerging compounds. Current non-targeted detection relies on high-resolution mass spectrometry, which has problems such as difficult result interpretation and lack of automation. A large number of unknown peaks still need to be manually analyzed. Small online monitoring mass spectrometers (such as portable mass spectrometers and ion mobility spectrometers (IMS)) can achieve rapid detection, but are limited by database coverage and algorithmic capabilities, making it difficult to accurately identify unknown pollutants in field conditions. Therefore, how to provide a method and system for unknown pollutant non-target identification based on mass spectrometry and generative model is a problem that needs to be solved by those skilled in the art. SUMMARY

[0003] Therefore, the present application provides a method and system for unknown pollutant non-target identification based on mass spectrometry and generative model, which realizes the full-process automation of pollutant structure and concentration from raw mass spectrometry data by combining standard library matching and intelligent generative algorithm.

[0004] To achieve the above purpose, the present application provides the following technical scheme: A method for unknown pollutant non-target identification based on mass spectrometry and generative model, comprising the following steps: S1, pretreating the water sample to be tested, and collecting mass spectrometry data of the water sample to be tested; S2, denoising, standardizing, and correcting the original mass spectrometry data to output a standardized feature for each precursor; S3, inputting the standardized feature into a pre-trained generative model to generate a plurality of candidate molecular formulas; S4, using chemical and physical constraints to exclude candidate molecular formulas that do not meet the rules; S5, performing a rule-driven algorithm to check the remaining candidate molecular formulas to obtain candidate pollutant molecular structures; S6, comprehensive scoring and ranking of candidate pollutant molecular structures by chemical priori and environmental priori; S7, semi-quantitative or relative quantitative concentration determination of the candidate pollutant molecular structures after scoring and ranking.

[0005] Optionally, S1 is specifically: The sample of the water body to be measured is subjected to automatic solid-phase extraction enrichment and concentration to facilitate the measurement of pollutants, and the volume and dilution / concentration multiple are recorded; liquid chromatography-high resolution mass spectrometry is used to separate and detect the sample of the water body to be measured, the mass spectrometry acquisition mode is selected to switch between positive and negative polarity to cover more types of organic pollutants; and a data non-dependent acquisition mode is used to obtain complete MS / MS fragment information, so as to ensure that low-abundance and unknown substances are not missed, obtain a full-scan mass spectrum, accurate mass-intensity information of all ions, obtain ion fragment spectrum by DIA technology, and further distinguish isomers by combining the collision cross section parameters of the compounds obtained synchronously in the acquisition process as quantifiable molecular structure characteristics.

[0006] Optionally, S2 is specifically: The original mass spectrometry data is sequentially subjected to peak intensity standardization and denoising, de-redundant peak and isotope recognition, charge and adduct correction, retention time correction, CCS standardization, and MS / MS spectrum normalization processing, and the standardized features are output: ; In the formula, is the mass-to-charge ratio of the precursor ion, is the best estimate of the charge number, is the adduct type, is the neutral molecular mass, is the corrected retention time, is the collision cross section, is the isotope distribution vector, is the normalized fragment spectrum vector, is the normalized intensity of the parent peak.

[0007] Optionally, the generated model in S3 is specifically: A conditional generation model is trained based on the standardized features, and a Transformer decoder architecture is used for training, the standardized features are input into a Transformer Encoder to obtain a conditional embedding c During training, the conditional embedding c is concatenated in front of the target molecular formula SMILES as a prefix token sequence, so that the generation model learns , i.e., given the conditional embedding cThe lower molecular formula is SMILES.

[0008] Optionally, S4 is specifically: For candidate molecular formula , the element composition is , and the unsaturation degree DBE is calculated: ; Wherein, X represents a halogen element, x is the value of X, the constraint DBE≥0, and the candidate molecular formula not meeting the constraint is removed.

[0009] Optionally, the rule-driven algorithm in S5 is specifically: Get common neutral loss fragments , for candidate molecular formula , calculate the parent ion neutral molecular mass and infer the loss common neutral molecules, find the corresponding fragment mass peaks in the fragment spectrum , if it meets: ; Then it is considered that there is neutral loss evidence in the fragment spectrum, and the sum of the weights of all matched neutral loss fragments is: ; Wherein, is the loss weight, is the loss threshold, and the sample mean is lower than 0.3, then it is removed; For each candidate molecular formula , the possible neutral loss is deduced through the fragmentation type, each neutral loss is regarded as a node, and the weight of the edge depends on whether the fragmentation type is reasonable in chemistry, which is defined as: ; In the formula, represents the weight of the edge e , the fragmentation probability is , the probability of common fragmentation is high, the weight score is low, the weight of uncommon fragmentation score is high, the node and edge form a fragmentation tree, which indicates how the candidate structure is split into fragments in the experimental spectrum, and whether the fragmentation tree can be connected to most of the fragments appearing in the experimental spectrum in a reasonable fragmentation way: ; In the formula, is the matching ratio, is the average mass error of all matching peaks, ​The maximum error allowed, less than 0.4 is rejected.

[0010] Optionally, S6 is specifically: The chemical structure is scored, and the scoring formula is: ChemPrior = -w1 SA_score - w2 Alerts - w3 complexcity; In the formula, ChemPrior is the chemical structure score, SA_score is the structure feasibility, Alerts is the structure risk alert number, and complexcity is the structure complexity, w1, w2, w3 are the weights of SA_score, Alerts, and complexcity respectively; A known pollutant library is established, and the maximum value of ECFP4 / Tanimoto of the candidate pollutant molecular structure and the library molecule is calculated: ; In the formula, Tanimoto is the similarity of two molecular fingerprints, is the similarity score, c is the current candidate pollutant molecular structure, d is the known pollutant structure, FP indicates the calculation of ECFP4; Calculate the substructure score : ; In the formula, is the substructure weight, is the first k characteristic fragment; The final score is defined as: ; In the formula, , , are the weights of ChemPrior, , respectively, and the five candidate pollutant molecular structures with the highest score are selected for concentration determination.

[0011] Optionally, S7 is specifically: Fit the log response factor model with a set of representative calibrators: ; In the formula, is the response factor, is the response factor model, a molecular descriptor calculated from the candidate contaminant molecular structure; predicting the unknown candidate contaminant molecular structure c in peak area calculating the estimated concentration : ; wherein, is the internal standard peak area in the same batch, is the total dilution factor before injection, is the pre-treatment recovery rate, is the post-pre-treatment dissolution volume, is the original sampling volume; analog-based surrogate quantification, if the most similar analog s is selected, then calculate: ; wherein, is the unknown concentration, is the internal standard normalized peak area of the unknown, is the internal standard normalized peak area of the similar known, is the ionization efficiency of the unknown, is the ionization efficiency of the known, is the known concentration of the known; calculate the confidence interval and uncertainty, combine the prediction variance of , peak area repeatability, recovery rate fluctuation, etc., to calculate the 95% interval of the concentration C .

[0012] ; wherein, represents the relative uncertainty of the concentration, represents the error of the peak area ratio, represents the uncertainty of the dilution factor, represents the uncertainty of the sample recovery rate, represents the uncertainty of the volume ratio, represents the uncertainty of the response factor prediction.

[0013] An unknown contaminant non-target identification system based on mass spectrometry and a generative model, applying the above unknown contaminant non-target identification method based on mass spectrometry and a generative model, comprising: a mass spectrometry acquisition module, configured to acquire mass spectrometry data of a water body sample to be tested; a data processing module, connected with the mass spectrometry acquisition module, configured to pre-process the original mass spectrometry data; The candidate generation module, connected to the data processing module, is used to generate and screen the molecular structures of candidate pollutants. The comprehensive scoring module, connected to the candidate generation module, is used to comprehensively evaluate and rank the molecular structures of candidate pollutants. The quantitative analysis module, connected to the comprehensive scoring module, is used to perform semi-quantitative or relative quantitative concentration determination of the molecular structure of candidate pollutants.

[0014] As can be seen from the above technical solutions, compared with the prior art, the present invention provides a method and system for the non-targeted identification of unknown pollutants based on mass spectrometry and generative models, which has the following beneficial effects: The present invention realizes full-process automation from data preprocessing to structure identification and then to concentration analysis, greatly reducing manual intervention and improving analytical throughput; The present invention combines rule-driven expert systems with data-driven generative models, which can effectively identify and analyze unknown pollutants that do not exist in the standard library; Based on the quantitative strategy of predicting response factors and analogs, valuable concentration information can still be obtained in the absence of standards, enhancing the practicality of the method. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0016] Figure 1 This is a flowchart of the method for non-targeted identification of unknown pollutants according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] This invention discloses a method for targetless identification of unknown pollutants based on mass spectrometry and a generative model, such as... Figure 1 As shown, it includes the following steps: S1. Pre-process the water sample to be tested and collect the mass spectrometry data of the water sample to be tested; S2. Denoise, standardize, and correct the raw mass spectrometry data, and output a standardized feature for each precursor. S3, input the normalized features into a pre-trained generative model to generate a plurality of candidate molecular formulas; S4, use chemical and physical constraints to exclude candidate molecular formulas that do not conform to the rules; S5, perform a rule-driven algorithm to check the remaining candidate molecular formulas to obtain candidate contaminant molecular structures; S6, comprehensively score and rank the candidate contaminant molecular structures by chemical and environmental priors; S7, perform semi-quantitative or relative quantitative concentration determination on the scored and ranked candidate contaminant molecular structures.

[0019] Further, S1 is specifically: The sample of the water body to be measured is automatically enriched and concentrated by solid phase extraction to facilitate the measurement of the pollutants, and the volume and dilution / concentration multiple are recorded; the water sample concentrate after pretreatment is quantitatively sampled by an automatic sampler, in this embodiment, the quantitative sampling volume is set to 5 μL, liquid chromatography-high resolution mass spectrometry (LC-HRMS) is used to separate and detect the water body sample to be measured, the mass spectrometry acquisition mode is selected to switch between positive and negative polarity to cover more types of organic pollutants; and a data-independent acquisition method is used to obtain complete MS / MS fragment information, so that low-abundance and unknown substances are not missed, a full scan mass spectrum is obtained, accurate mass-intensity information of all ions, all ions are fragmented by DIA technology to obtain ion fragment spectrum; and the collision cross-section area parameters of the compounds obtained synchronously during the acquisition process are combined as quantifiable molecular structure characteristics to further distinguish isomers.

[0020] In the embodiment of the present application, if the initial volume is concentrated to after solid phase extraction and concentration, the concentration multiple is: ; In the embodiment of the present application, the system is calibrated and quality controlled before detection to ensure that the subsequent concentration determination is reliable, including: 1. Internal standard addition to eliminate instrument drift: a stable isotope internal standard is recommended, which is uniformly added to all samples, blanks and calibration solutions before sampling The peak area of the target substance is recorded as , the peak area of the internal standard is recorded as , and the response ratio is defined as: ; 2. Establish a standard curve to determine the concentration by signal: a regression is established between the concentration C and the response ratio R: ; 3. Determine the limit of detection (LOD) and limit of quantification (LOQ) by measuring the noise standard deviation of the blank (water without pollutants) , combined with the calibration curve slope k, to distinguish whether the signal is noise: ; ; 4. Control the quality error: compare the molecular weight measured by the instrument with the theoretical value , calculate the relative error , and require that the mass remains : .

[0021] Further, S2 is specifically: The original mass spectrum data is sequentially subjected to peak intensity standardization and denoising, de-redundant peak and isotope recognition, charge and adduct correction, retention time correction, CCS standardization, MS / MS spectrum normalization processing, and output standardized features: ; In the formula, is the mass-to-charge ratio of the precursor ion, is the optimal estimate of the charge number, is the adduct type, is the neutral molecular mass, is the corrected retention time, is the collision cross section area, is the isotope distribution vector, is the normalized fragment spectrum vector, is the normalized intensity of the parent peak.

[0022] In the embodiments of the present application, peak intensity standardization and denoising include: Total ion flow TIC normalization: ; ; Divide the intensity of each peak by the total intensity of this time , different sample amounts and spray states will make the overall intensity fluctuate; after normalization, the relative strengths of each peak can be compared; To avoid noise interference, only peaks higher than the noise are retained to reduce false positives in subsequent algorithms, and the signal-to-noise ratio threshold is defined as: ; In the formula, h is the peak height, For noise standard deviation, define detection peak , quantitative peak ; De-redundant peaks and isotopic recognition includes: Recognize isotopic peaks, identify parent peak M and isotopic peaks M+1, M+2, isotopic spacing And charge Coarse judgment: ; ; The theoretical relative abundance of isotopes is predictable, compare the observed isotopic intensity with the predicted intensity according to the natural abundance, and use the cosine similarity: ; Similarity, Observed isotopic intensity, Predicted intensity according to natural abundance, Highly accurate is indeed a cluster of isotopic peaks; Charge and adduct correction includes: The molecules in the mass spectrum may have different charge numbers z, and may form different adduct ions ([M+H]+, [M+Na]+, [M-H]-, etc.), from the observed charged / adduct ions, restore the true molecular mass without adduction, no charge. For the observed , z and candidate adduct The neutral molecular mass is: ; In the formula, Observed mass-to-charge ratio, Charge number, Mass of adduct, Corrected neutral molecular mass; Test whether the selected z matches the entire cluster of isotopic spacings to avoid a series of errors caused by incorrect charge: ; In the formula, Spacing threshold.

[0023] For all possible charge / adduct combinations (z, A), fit the mass error and isotopes into a target, and convert the difference between measurement and theory into a similarity score of 0-1. The smaller the error, the more reliable; Numerical implementation through Gaussian kernel: ; ; In the formula, similarity score, mass error, theoretical mass-to-charge ratio; choose the most reasonable one from all possible charge / adduct combinations: ; where, , , are the weights of , , adduct prior (less common adducts get lower score), considering [M+H] + is more common than [H+K] + in the real world, incorporate this experience as prior into the result to get ; Retention time correction includes: considering that the same compound in different batches may peak earlier / later by a few seconds, mispairing will occur if not aligned, therefore stretch / translate the retention time RT coordinates of this batch to the reference coordinates, correct the systematic drift with linear correction or nonlinear fitting: ; Gaussian similarity can be used to compare the consistency of observed RT and standard RT when matching: ; where, is the time difference between observed and standard values, is the allowed error; CCS standardization includes: Even if the mass is the same, the molecule may have different conformations; CCS is a key evidence to distinguish the look-alikes, compare the molecular shape indicators to match, compare the observed values with the reference values: ; ; where, is the matching score, is the allowed CCS error, is the observed value, is the reference value, is the difference; MS / MS spectrum normalization processing includes: ​Different instruments and batches may exhibit slight fluctuations in m / z; binning facilitates comparison; L2 normalization and cosine similarity, unaffected by overall strength, aggregate sparse, tolerance-controlled fragment spectra into fixed m / z bins, perform length normalization, tolerance alignment, and binning to facilitate vector comparison. ; ; In the formula, For the normalized fragment spectral vector, It is the L2 norm, that is , This is the bin strength vector obtained in the previous step. This represents the intensity of the corresponding ion peak. This represents the mass-to-charge ratio of the corresponding ion peak. For the first k Each bin center is a fixed m / z position after the spectrum is discretized. The allowable quality deviation tolerance range, For the first The strength value of a standardized quality window (or sub-box).

[0024] Furthermore, the generative model in S3 is specifically as follows: A conditional generative model is trained based on standardized features, using a Transformer decoder architecture. The standardized features are processed by a Transformer Encoder. Obtain conditional embedding c Conditional embedding during training c As a prefix token sequence, it is appended to the target molecular formula SMILES, enabling the generative model to learn. That is, given conditional embedding c The molecular formula SMILES is generated.

[0025] In this embodiment of the invention, the specific process of the training phase is as follows: For mixed inputs of different spectral feature dimensions, the localization is: c = [continuous features] || [discrete feature embedding] || [vector features], mapping the spectral features into an embedding vector: ;SMILES sequence labeled as ; each of them It is a token, such as an atom or bond, that embeds conditions into h. c As a virtual token, it is appended to the beginning of the input sequence: [s0] = [COND], embedding(s0)=h c ; The model input becomes: [h c , embedding(s1), embedding(s2), ... , embedding(s T-1 )]; Transformer decoder learns autoregressive conditional probabilities: ; The training objective function is defined as maximizing the conditional likelihood: ; The specific process of the reasoning generation stage is as follows: Computational Conditional Embedding Starting with the [COND] token, SMILES are generated autoregressively: ; After inputting the standardized features obtained from the instrument, several candidate SMILES are generated.

[0026] Furthermore, S4 specifically refers to: S4 specifically refers to: For candidate molecular formulas The elemental composition is Calculate the degree of unsaturation (DBE): ; Where X represents a halogen element. x The value of X is used to constrain DBE≥0, and candidate molecular formulas that do not meet the constraints are eliminated; if DBE<0, it means that the number of valence bonds in the molecule does not match the number of atoms (for example, too many hydrogens or unbalanced atomic valences), and such a structure is chemically impossible.

[0027] Furthermore, the rule-driven algorithm in S5 is specifically as follows: Obtain common neutral loss fragments (e.g., H₂O 18.0106, CO₂ 43.9898, SO₃ 79.9568, HF₂ 0.0062), for candidate molecular formulas Calculate mother ionic neutral molecular weight It also inferred the loss of common neutral molecules and searched for corresponding fragment mass peaks in the fragment spectra. If the following conditions are met: ; Therefore, it is assumed that there is evidence of neutral loss in the fragment spectrum, which is all the matching neutral loss fragments. Weighted summation: ; in, For loss weights, the commonly used loss weights are higher. The loss threshold, Samples with a value less than 0.3 of the overall sample mean were removed. To test whether a generated molecule can naturally explain the observed fragment spectrum, for each candidate molecular formula... Possible neutral losses are derived from the fracture types. Each neutral loss is treated as a node, and the weight of an edge depends on whether the fracture type is chemically reasonable, defined as follows: ; In the formula, Representing an edge e Weight, This represents the probability of fracture, specifically the probability of common fractures. High score, high weight Low score, uncommon fracture weight A high score indicates that the nodes and edges form a fragmented tree, representing how the candidate structure splits into fragments in the experimental spectrum. The calculation determines whether the fragmented tree can connect to most of the fragments appearing in the experimental spectrum in a reasonable breaking manner. ; In the formula, For matching ratio, The average quality error of all matched peaks, To represent the maximum permissible error, in this embodiment, 10 ppm is used. If the value is less than 0.4, it will be discarded.

[0028] Furthermore, S6 specifically refers to: Higher scores are given to chemically more common, rational, and easily synthesized structures, while absurd, unstable, or overly complex structures are downgraded or eliminated; RDkits are used to examine candidate molecules for valence, charge, and aromaticity; PAINS and unstable SMARTS groups, such as acyl halides, hypochlorites, active enones, and nitrosamines, are searched; the number of heavy atoms (NHA), ring coefficient (Nring), number of fused / spirocyclic rings (Nfuesd), number of chiral centers (Nchiral), and number of rotatable bonds (Nrot) are detected; the SA score is calculated, with higher scores indicating greater difficulty in synthesis; and the chemical structure is scored using the following formula: ChemPrior = -w1 SA_score - w2 Alerts - w3 complexcity; In the formula, ChemPrior is the chemical structure score, SA_score is the structural feasibility, Alerts is the number of structural risk warnings, complexity is the structural complexity, and w1, w2, and w3 are the weights of SA_score, Alerts, and complexity, respectively. Establish a library of known pollutants (pesticides, drugs, PFAS, surfactants, dyes, etc.), and calculate the maximum value of ECFP4 / Tanimoto relative to the library molecules for the molecular structures of candidate pollutants. ; In the formula, Tanimoto represents the similarity between the two molecular fingerprints. To score the similarity, c The molecular structures of the current candidate pollutants, d Given the known structure of the pollutants, FP This indicates the calculation of ECFP4; in this embodiment, EnvSim > 0.6 is considered "family similar", and < 0.3 is considered unrelated; Without simulating fragmentation of the candidate structure, check the observed spectra for typical "diagnostic fragments," such as characteristic fragments. (e.g., benzene rings, quaternary ammonium ions, etc.) Search for specific m / z values ​​in the fragment spectrum. For example, the diagnostic peak for the benzene ring is at m / z 77±0.3. Check if a diagnostic peak exists. If detected, add points. A higher score indicates more evidence supporting the presence of these key substructures in the molecule; substructure scores are calculated. : ; In the formula, Assigning weights to substructure features: common features such as CH2 and H2O are assigned low weights, while less common but highly diagnostic fragments such as halobenzene ions and acyl ions are assigned high weights. For the first k One feature fragment; The final score is defined as follows: ; In the formula, , , They are ChemPrior, , The weights were determined, and the molecular structures of the five candidate pollutants with the highest scores were selected for concentration determination.

[0029] Furthermore, S7 specifically refers to: Response factor prediction based on ionization efficiency (IE) is used to estimate the area per unit concentration, and a log response factor model is fitted using a set of representative calibrators: ; In the formula, As the response factor, In this embodiment, the model can be an SVR / NN, which is a response factor model. These are molecular descriptors calculated from the molecular structures of candidate pollutants; Prediction of molecular structure of unknown candidate pollutants With peak area Calculate the estimated concentration : ; In the formula, This refers to the area of ​​the standard peak within the same batch. This is the overall dilution factor before injection. For pretreatment recovery rate, This represents the volume of solution after pretreatment. This represents the original sampling volume; Based on the quantification of analogues, if the most structurally similar analogue *s* can be selected, then the following calculation is performed: ; In the formula, For the concentration of an unknown substance, The internal standard normalized peak area of ​​the unknown substance. The internal standard normalized peak area for similar known objects. This is the overall dilution factor before injection. For pretreatment recovery rate, The ionization efficiency of the unknown substance. Given the ionization efficiency of the substance, The known concentration of the known substance; in embodiments of the invention, if the analogy response is used directly ( ≈ In the above formula, the IE ratio can be set to 1, and the deviation is included in the uncertainty. Calculate the confidence interval and uncertainty, and put The prediction variance, peak area repeatability, and recovery rate fluctuations are combined to calculate the concentration. C The 95% range: ; In the formula, Represents the relative uncertainty of concentration. The error representing the peak area ratio is, in this embodiment, taken as an empirical value of 3%. This represents the uncertainty of the dilution factor; in this embodiment, the empirical value is taken as 2%. The uncertainty in the sample recovery rate. The uncertainty representing the volume ratio is, in this embodiment, taken as an empirical value of 1.5%. This represents the uncertainty of the response factor prediction. In this embodiment, the empirical value is 7% for experimental calibration and 20% for the RF predicted by the machine learning model.

[0030] and Figure 1 Corresponding to the method described above, this invention also discloses a non-targeted identification system for unknown pollutants based on mass spectrometry and a generative model. The application of the aforementioned non-targeted identification method for unknown pollutants based on mass spectrometry and a generative model includes: The mass spectrometry acquisition module is used to acquire mass spectrometry data from the water sample to be tested. The data processing module, connected to the mass spectrometry acquisition module, is used to preprocess the raw mass spectrometry data; The candidate generation module, connected to the data processing module, is used to generate and screen the molecular structures of candidate pollutants. The comprehensive scoring module, connected to the candidate generation module, is used to comprehensively evaluate and rank the molecular structures of candidate pollutants. The quantitative analysis module, connected to the comprehensive scoring module, is used to perform semi-quantitative or relative quantitative concentration determination of the molecular structure of candidate pollutants.

[0031] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0032] Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for targetless identification of unknown pollutants based on mass spectrometry and generative models, characterized in that, Includes the following steps: S1. Pre-process the water sample to be tested and collect the mass spectrometry data of the water sample to be tested; S2. Denoise, standardize, and correct the raw mass spectrometry data, and output a standardized feature for each precursor. S3. Input the standardized features into the pre-trained generative model to generate multiple candidate molecular formulas; S4. Use chemical and physical constraints to eliminate candidate molecular formulas that do not conform to the rules; S5. Execute the rule-driven algorithm to check the remaining candidate molecular formulas and obtain the molecular structures of candidate pollutants; S6. The molecular structures of candidate pollutants are comprehensively scored and ranked based on chemical and environmental priors. S7. Perform semi-quantitative or relative quantitative concentration determination of the molecular structure of candidate pollutants after scoring and ranking.

2. The method for targetless identification of unknown pollutants based on mass spectrometry and a generative model according to claim 1, characterized in that, S1 specifically refers to: Automated solid-phase extraction (SPE) enrichment and concentration of the water samples were performed to facilitate the detection of pollutants, and the volume and dilution / concentration factor were recorded. Liquid chromatography-high resolution mass spectrometry (LC-MS-MS) was used to separate and detect the water samples, with positive / negative polarity switching of the mass spectrometry acquisition mode to cover more types of organic pollutants. A data-independent acquisition method was used to obtain complete MS / MS fragment information, ensuring that low abundance and unknown substances were not missed, resulting in a full-scan mass spectrum with accurate mass-intensity information for all ions. DIA technology was used to fragment all ions to obtain ion fragmentation maps. By combining the collision cross-sectional area parameters of the compounds obtained simultaneously during the acquisition process, as a quantifiable molecular structural feature, isomers can be further distinguished.

3. The method for targetless identification of unknown pollutants based on mass spectrometry and a generative model according to claim 1, characterized in that, S2 specifically refers to: The raw mass spectrometry data were sequentially processed with peak intensity normalization and denoising, redundancy peak removal and isotope identification, charge and adduct correction, retention time correction, CCS normalization, and MS / MS spectrum normalization, outputting normalized features: ； In the formula, The mass-to-charge ratio of the precursor ion. For the best estimate of the charge number, It is an adduct type. Neutral molecular weight This is the corrected retention time. For the collision cross-sectional area, This is the isotope distribution vector. For the normalized fragment spectral vector, This represents the normalized intensity of the parent peak.

4. The method for targetless identification of unknown pollutants based on mass spectrometry and a generative model according to claim 1, characterized in that, The generative model in S3 is specifically as follows: A conditional generative model is trained based on standardized features, using a Transformer decoder architecture. The standardized features are processed by a Transformer Encoder. Obtain conditional embedding c Conditional embedding during training c As a prefix token sequence, it is appended to the target molecular formula SMILES, enabling the generative model to learn. That is, given conditional embedding c The molecular formula SMILES is generated.

5. The method for targetless identification of unknown pollutants based on mass spectrometry and a generative model according to claim 1, characterized in that, S4 specifically refers to: For candidate molecular formulas The elemental composition is Calculate the degree of unsaturation DBE: ； Where X represents a halogen element. x Given the value of X, constrain DBE≥0, and eliminate candidate molecular formulas that do not meet the constraints.

6. The method for targetless identification of unknown pollutants based on mass spectrometry and a generative model according to claim 1, characterized in that, The rule-driven algorithm in S5 is as follows: Obtain common neutral loss fragments For candidate molecular formulas Calculate mother ionic neutral molecular weight It also inferred the loss of common neutral molecules and searched for corresponding fragment mass peaks in the fragment spectra. If the following conditions are met: ； Therefore, it is assumed that there is evidence of neutral loss in the fragment spectrum, which is all the matching neutral loss fragments. Weighted summation: ； in, To lose weight, The loss threshold, Samples with a value less than 0.3 of the overall sample mean were removed. For each candidate molecular formula Possible neutral losses are derived from the fracture types. Each neutral loss is treated as a node, and the weight of an edge depends on whether the fracture type is chemically reasonable, defined as follows: ； In the formula, Representing an edge e Weight, This represents the probability of fracture, specifically the probability of common fractures. High, weight Low score, uncommon fracture weight A high score indicates that the nodes and edges form a fragmented tree, representing how the candidate structure splits into fragments in the experimental spectrum. The calculation determines whether the fragmented tree can connect to most of the fragments appearing in the experimental spectrum in a reasonable breaking manner. ； In the formula, For matching ratio, The average quality error of all matched peaks, For the maximum allowable error, If the value is less than 0.4, it will be discarded.

7. The method for targetless identification of unknown pollutants based on mass spectrometry and a generative model according to claim 1, characterized in that, S6 specifically refers to: The chemical structure is scored using the following formula: ChemPrior = -w1 SA_score - w2 Alerts - w3 complexcity; In the formula, ChemPrior is the chemical structure score, SA_score is the structural feasibility, Alerts is the number of structural risk warnings, complexity is the structural complexity, and w1, w2, and w3 are the weights of SA_score, Alerts, and complexity, respectively. Establish a known pollutant library and calculate the maximum value of ECFP4 / Tanimoto relative to the library molecules for candidate pollutant molecular structures: ； In the formula, Tanimoto represents the similarity between the two molecular fingerprints. To score the similarity, c The molecular structures of the current candidate pollutants, d Given the known structure of the pollutants, FP This indicates the calculation of ECFP4; Calculate substructure score : ； In the formula, For substructure weights, For the first k One feature fragment; The final score is defined as follows: ； In the formula, , , They are ChemPrior, , The weights were determined, and the molecular structures of the five candidate pollutants with the highest scores were selected for concentration determination.

8. The method for targetless identification of unknown pollutants based on mass spectrometry and a generative model according to claim 1, characterized in that, S7 specifically refers to: Fit the log response factor model using a set of representative calibrators: ； In the formula, As the response factor, For the response factor model, These are molecular descriptors calculated from the molecular structures of candidate pollutants; Prediction of molecular structure of unknown candidate pollutants With peak area Calculate the estimated concentration : ； In the formula, This refers to the area of ​​the standard peak within the same batch. This is the overall dilution factor before injection. For pretreatment recovery rate, This represents the volume of solution after pretreatment. This represents the original sampling volume; Based on the quantification of analogues, if the most structurally similar analogue *s* can be selected, then the following calculation is performed: ； In the formula, For the concentration of an unknown substance, The internal standard normalized peak area of ​​the unknown substance. The internal standard normalized peak area for similar known objects. The ionization efficiency of the unknown substance. Given the ionization efficiency of the substance, The known concentration of a known substance; Calculate the confidence interval and uncertainty, and put The prediction variance, peak area repeatability, and recovery rate fluctuations are combined to calculate the concentration. C The 95% range. ； In the formula, Indicates the relative uncertainty of concentration. This indicates the error in the peak area ratio. This indicates the uncertainty of the dilution factor. The uncertainty in the sample recovery rate. The uncertainty in volume ratio is expressed as... This represents the uncertainty in the prediction of the response factor.

9. A targetless identification system for unknown pollutants based on mass spectrometry and a generative model, characterized in that, The method for targetless identification of unknown pollutants based on mass spectrometry and a generative model, according to any one of claims 1-8, comprises: The mass spectrometry acquisition module is used to acquire mass spectrometry data from the water sample to be tested. The data processing module, connected to the mass spectrometry acquisition module, is used to preprocess the raw mass spectrometry data; The candidate generation module, connected to the data processing module, is used to generate and screen the molecular structures of candidate pollutants. The comprehensive scoring module, connected to the candidate generation module, is used to comprehensively evaluate and rank the molecular structures of candidate pollutants. The quantitative analysis module, connected to the comprehensive scoring module, is used to perform semi-quantitative or relative quantitative concentration determination of the molecular structure of candidate pollutants.

Images