A method for non-targeted screening and identification of compound concentrations based on machine learning

Abstract

The present invention relates to the field of new pollution monitoring technology, and in particular to a method for non-targeted screening and identification of compound concentrations based on machine learning. Based on high-resolution mass spectrometry information, fingerprint information and chemical descriptors of compounds, the present invention adopts neural networks and Bayesian statistical methods to construct a method for non-targeted screening and identification of compound concentrations based on machine learning. The problem that the concentration of compounds cannot be quantitatively obtained in the absence of standard substances is solved; in the case of standard substances, the quantitative acquisition of concentration cannot be completed in the non-targeted screening stage, and liquid chromatography-triple quadrupole mass spectrometry needs to be used again for quantification, resulting in the problem of complicated and time-consuming procedures.

Description

Technical Field

[0001] The present invention relates to the field of new pollution monitoring technology, and in particular to a method for non-targeted screening and identifying compound concentrations based on machine learning. Background Art

[0002] New pollutants are a general term for a class of chemical substances that can be detected in the environment and can pose significant risks and hidden dangers to human health and environmental safety even at low doses. They mainly include tens of thousands of chemical substances such as persistent organic pollutants, endocrine disruptors, antibiotics and microplastics. Quantitative analysis of new pollutants is an important basis for environmental risk management.

[0003] In the existing technology, the quantitative method of new pollutants must rely on existing standards, and use chromatography, mass spectrometry, spectroscopy and other analytical techniques to separate and detect the target compounds in the sample. Standards are crucial for drawing standard curves because they provide a benchmark of known concentration, allowing the measured response signal to be compared with the known concentration during the analysis process, thereby accurately determining the concentration of new pollutants in environmental samples. However, for tens of thousands of compounds, not every compound has a corresponding standard, which greatly increases the complexity and difficulty of quantitative analysis of new pollutants. Therefore, the development of efficient alternative analytical methods has become an important direction to address this challenge in order to conduct more comprehensive environmental monitoring and risk assessment. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the purpose of the present invention is to provide a method for non-targeted screening and identification of compound concentrations based on machine learning, aiming to solve two problems: one is the problem that the concentration of the compound cannot be quantitatively obtained in the absence of a standard; the other is that in the presence of a standard, the concentration cannot be quantitatively obtained in the non-targeted screening stage, and liquid chromatography-triple quadrupole mass spectrometry needs to be used again for quantification, resulting in a complicated and time-consuming procedure.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0006] A method for non-targeted screening and identifying compound concentrations based on machine learning is provided, comprising the following steps:

[0007] S1. Constructing a trained deep residual neural network for predicting concentration, and obtaining the compound structure, mass spectrum fusion coding, predicted concentration and its predicted relative deviation during the training of the deep residual neural network;

[0008] S2. constructing a trained compound prediction accuracy discrimination model, and obtaining a reconstruction error of the compound prediction accuracy discrimination model during the training process;

[0009] S3, obtaining the structural information of the predicted compound and the internal standard compound of the spiked sample: molecular descriptor, MACCS molecular fingerprint spectrum, Estate1 molecular fingerprint spectrum and mass spectrum information: peak area;

[0010] S4, using the structural information and mass spectrum information of the predicted compound in the spiked sample and the internal standard compound as inputs to the trained deep residual neural network to obtain the compound structure, mass spectrum fusion code and predicted concentration of the predicted compound in the spiked sample and the internal standard compound;

[0011] S5. Using the predicted concentration output of the trained deep residual neural network as the first concentration of the compound to be predicted; using the compound structure and mass spectrum fusion encoding of the spiked compound to be predicted and the internal standard compound as the input of the trained compound prediction accuracy discrimination model, and judging the accuracy of the concentration of the compound to be predicted according to the reconstruction error of the compound prediction accuracy discrimination model during the training process;

[0012] S6, obtaining an average relative response factor of the compound to be predicted relative to the internal standard compound according to the first concentration and peak area of ​​the compound to be predicted in the spiked sample and the first concentration and peak area of ​​the internal standard compound;

[0013] S7, calculating the second concentration of the compound to be predicted based on the peak area and average relative response factor of the compound to be predicted in the spiked sample, the added amount and peak area of ​​the internal standard compound;

[0014] S8, obtaining a relative deviation between the second concentration of the compound to be predicted and the relative spiked amount according to the added amount of the compound to be predicted and the second concentration of the compound to be predicted;

[0015] S9. According to the accuracy information in S5, the relative deviation of the deep residual neural network model predicting a single compound during training is used as prior information for high-precision and low-precision compounds, and the relative deviation of the second concentration and the relative spiked amount of the compound to be predicted is used as posterior information. The lognormal distribution is used as the prior distribution and the normal distribution is used as the posterior distribution. The prediction relative deviations of high-precision and low-precision compounds are calculated by the Bayesian method respectively;

[0016] S10. Correcting the second concentration of the compound to be predicted by predicting the relative deviation to obtain the final concentration and range of the compound to be predicted.

[0017] Furthermore, in S1, the method for constructing a deep residual neural network is specifically as follows:

[0018] Compound features and peak areas are used as input, and concentration is used as output; a residual neural network model is built based on pytorch; MACCS molecular fingerprint spectra and Estate1 molecular fingerprint spectra are used as discrete feature inputs, and molecular descriptors are used as continuous feature inputs. After passing through one linear layer, they are fused into at least 6 subsequent layers of neural networks in the model, and then the peak area is input. Finally, the compound's escape features are output after passing through at least 2 layers of neural networks, and finally, the compound's concentration is output after passing through at least 2 layers of neural networks.

[0019] Furthermore, in S1, the relative deviation of the model prediction for a single compound The calculation formula is:

[0020]

[0021] Where: is the true concentration; is the predicted concentration; is the number of predicted concentrations of the same compound.

[0022] Furthermore, the method of S6 is specifically as follows:

[0023] Obtain at least 10 peak area-concentration pairs of the compound to be predicted and the internal standard compound, perform Monte Carlo simulation at least 50,000 times, and obtain the relative response factors of at least 5 peak area-concentration pairs each time RF Calculate and obtain the average relative response factor of the predicted compound relative to the internal standard compound ;

[0024]

[0025]

[0026] Where: is the relative response factor, dimensionless;

[0027] : concentration of internal standard compound, μg / L;

[0028] : The first concentration of the compound to be predicted in the spiked sample, μg / L;

[0029] : Peak area of ​​internal standard compound;

[0030] : Peak area of ​​the compound to be predicted;

[0031] : average relative response factor, dimensionless;

[0032] n: number of series points, n≥5.

[0033] Furthermore, the methods of S7 and S8 are specifically as follows:

[0034] The compound to be predicted is set to multiple concentration gradients and added to pure water or other environmental samples respectively, with multiple groups of parallels set for each concentration gradient; after pre-treatment, the peak area is obtained by high-resolution mass spectrometry, and the prediction calculation is performed according to the following formula to obtain the second concentration of the compound to be predicted and the relative deviation of the relative spiked amount ;

[0035]

[0036] Where: is the second concentration of the compound to be predicted, μg / L;

[0037] A: Peak area of ​​the compound to be predicted in the sample;

[0038] : Peak area of ​​internal standard compound in sample;

[0039] : The amount of internal standard compound added to the sample;

[0040]

[0041] Where: is the spike concentration; is the predicted concentration; is the number of spiked samples with the same spike concentration.

[0042] Furthermore, in S9, the prior distribution of the relative deviations of high-precision and low-precision compounds is as follows:

[0043]

[0044]

[0045]

[0046] Where:

[0047] rd: relative deviation probability distribution of high-precision or low-precision compounds, i.e., the prior distribution of relative deviation;

[0048] Log-Normal: log-normal distribution;

[0049] : estimated value of the rd mean;

[0050] : rd is the estimated value of the standard deviation;

[0051] : represents the average relative deviation of the i-th compound;

[0052] : the number of compounds;

[0053] Let the error event introduced by considering the instrument response and the pre-processing process be X, and let the relative deviation x obtained by S6 be the observed value of X;

[0054]

[0055] : The mean of the observed values ​​x, i.e. the prior of the relative deviation to be predicted;

[0056] Normal: normal distribution;

[0057] : The standard deviation of the observed value x, that is, the prior of the standard deviation to be predicted;

[0058] The conditional distribution of rd, that is, the posterior distribution of rd, is shown as follows:

[0059]

[0060] Where:

[0061] : The probability of relative deviation of the predicted concentration given X;

[0062] : is the total probability of rd, that is, the probability of the relative deviation in S1;

[0063] : is the joint probability distribution of X and rd, that is, the marginal probability of rd under the condition of X;

[0064] According to the observed value x, the Bayesian method MCMC is used to calculate based on the pymc computing package , obtain the posterior probability distribution of the mean relative deviation of the predicted concentration, with its mean as the relative deviation from the predicted concentration.

[0065] Further, in S10, the final concentration range of the compound to be predicted is obtained according to the following formula;

[0066]

[0067] Where:

[0068] : is the first concentration of the compound to be predicted;

[0069] : is the relative deviation of the predicted concentration.

[0070] The beneficial effects of the present invention are:

[0071] The method of the present invention comprehensively considers the high-resolution mass spectrum information, fingerprint information and chemical descriptors of the compound, and makes full use of neural networks and Bayesian statistical methods to effectively correct the errors caused by instrument response changes and pre-treatment processes, thereby maximizing the accuracy and reliability of concentration prediction. At the same time, an accuracy discrimination model is proposed, which provides a basis for subsequent optimization of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 It is the workflow diagram of the method of the present invention;

[0073] Figure 2 This is the architecture diagram of the compound concentration prediction model and the accuracy discrimination model;

[0074] Figure 3 Estimation framework diagram for quantitative relative deviation;

[0075] Figure 4 It is a schematic diagram of the high-resolution mass spectrometry peak area information and the quantitative results of liquid chromatography-triple quadrupole mass spectrometry of the compounds in the application case; the red dots represent the quantitative results of liquid chromatography-triple quadrupole mass spectrometry, the blue dots represent the predicted concentrations, and the blue lines represent the upper and lower limits of the predictions; the red marks show the relative deviation of the predicted concentrations and the prediction accuracy. DETAILED DESCRIPTION

[0076] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0077] Example

[0078] Reference Figure 1 , provides a method for non-targeted screening and identifying compound concentrations based on machine learning, which specifically includes the following steps:

[0079] 1. Dataset

[0080] There are two mixed standards, mixed standard 1 contains 67 compounds and mixed standard 2 contains 496 compounds. Mixed standard 1 is configured with 6 concentration points of 10, 20, 100, 200, 400, and 1000 μg / L, and mixed standard 2 is configured with 6 concentration points of 5, 10, 50, 100, 200, and 500 μg / L. Without pretreatment, the peak area of ​​each substance at each concentration point is directly measured by high-resolution mass spectrometry, and its calibration curve is obtained. After constructing the compound feature data, a data set is formed. The compound features include molecular descriptors (MolecularDescriptors), MACCS molecular fingerprints (MACCS fingerprints), Estate1 molecular fingerprints (EStatefingerprints), and peak area. Concentration-peak area data are generated according to the calibration curve, and 5% Gaussian noise is randomly added to the peak area. The sample set is randomly shuffled at a ratio of 60%, 20%, and 20%, and divided into a training set, a validation set, and a test set.

[0081] 2. Building a deep residual neural network

[0082] The compound features and peak area are used as input, and the concentration is used as output. A residual neural network model is built based on pytorch. MACCS molecular fingerprints (MACCS fingerprints) and Estate1 molecular fingerprints (EStatefingerprints) are used as discrete feature inputs, and molecular descriptors (Molecular Descriptors) are used as continuous feature inputs. After passing through a linear layer, they are fused into the model's subsequent 12-layer neural network, and then the peak area is input. Finally, the compound's escape features are output through a 3-layer neural network, and finally, the compound's concentration is output through a 2-layer neural network. The model loss function is SmoothL1Loss, the optimizer is Adam, and the flat area learning rate decay method is used to dynamically adjust the learning rate algorithm. The determination coefficient R 2 And absolute deviation (MAE) as model evaluation parameters, the results are as follows:

[0083]

[0084] Relative deviation statistics of model prediction results. Based on the model data set, the distribution of the average relative deviation between each predicted concentration and the actual concentration was analyzed. The relative deviation calculation formula is as follows, and the results are shown in the following table. The relative deviation of 95% of the compounds in the model prediction results is less than 30%.

[0085]

[0086] Where:

[0087] is the true concentration;

[0088] is the predicted concentration;

[0089] is the number of predicted concentrations of the same compound.

[0090]

[0091] 3. Construction and training of prediction accuracy discrimination model (i.e. model in S2)

[0092] A neural network based on an autoencoder was constructed. The data with a relative deviation less than or equal to 30% was used as the data set. The test set and training set were divided into 7:3. The escape output of the predicted detection model was used as the input of the model to obtain the reconstruction error of the model.

[0093] The 95th percentile of the minimum reconstructed absolute error of each compound at each concentration is used as the discrimination limit. Compounds with a reconstruction error greater than the above limit are defined as compounds with high prediction accuracy, and compounds with a limit lower than the above limit are defined as compounds with low prediction accuracy. The discriminator's discrimination results for the compounds in step 2 are shown in the following table. Among the compounds in the table with a relative deviation greater than 100%, 19 of them were identified as low-precision compounds after being discriminated by the discriminator. Among the compounds with a relative deviation of 60-100%, 8 of them were identified as low-precision compounds after being discriminated by the discriminator. Compounds with a relative deviation of less than 60% are basically identified as high-precision compounds.

[0094]

[0095] 4. Relative response factor calculation

[0096] According to the target compound, a suitable internal standard substance is selected. In this example, 38 target compounds and 1 internal standard compound are selected. The prediction accuracy information of the 38 compounds is obtained after the discriminant model is used, where 0 indicates high prediction accuracy and 1 indicates low prediction accuracy. The results are shown in the following table.

[0097]

[0098] Monte Carlo simulation was performed 500,000 times, and the relative response factors of the target compound and the internal standard compound were obtained at 6 concentration points each time. RF ), calculate and obtain the average relative response factor of the target compound ( ).

[0099]

[0100]

[0101] Where: is the relative response factor, dimensionless;

[0102] : internal standard compound concentration, μg / L;

[0103] : Target compound concentration, μg / L;

[0104] : Peak area of ​​internal standard compound;

[0105] : peak area of ​​target compound;

[0106] : average relative response factor, dimensionless;

[0107] n: number of series points, n≥5.

[0108] The calculation results are shown in the following table.

[0109]

[0110] 5. Forecast Error Estimation

[0111] The 38 target compounds were set up with two concentration gradients of 100 μg / L and 10 μg / L, and added to environmental samples and pure water samples, respectively. Four groups of parallel experiments were set up for each concentration gradient. 100 μg / L of internal standard compound was added to each matrix spiked sample. After pretreatment, the peak area was obtained by high-resolution mass spectrometry. The predicted concentration and relative deviation of the compound were obtained by prediction calculation according to the following formula.

[0112]

[0113] Where: is the concentration of the target compound in μg / L; A is the peak area of ​​the target compound in the sample; is the internal standard peak area in the sample; is the amount of internal standard added to the sample.

[0114]

[0115] Where: is the added amount; is the predicted concentration; is the number of spiked samples with the same spiked concentration, here it is 4.

[0116] The relative deviation obtained in step 2 is used as the prior information, and the relative deviation of the predicted concentration of the above samples is used as the posterior information based on the instrument response and pre-processing processes. It is assumed that the prior distribution is lognormal distribution and the posterior distribution is normal distribution.

[0117] Estimate the posterior distribution parameters of the relative deviation of the compound with high precision and low precision.

[0118]

[0119]

[0120]

[0121] Where:

[0122] rd: relative deviation probability distribution of high-precision or low-precision compounds, i.e., the prior distribution of relative deviation;

[0123] Log-Normal: log-normal distribution;

[0124] : estimated value of the rd mean;

[0125] : rd is the estimated value of the standard deviation;

[0126] : represents the average relative deviation of the i-th compound;

[0127] : the number of compounds;

[0128] Let the error event introduced by considering the instrument response and the pre-processing process be X, and let the relative deviation x obtained by S6 be the observed value of X;

[0129]

[0130] : The mean of the observed values ​​x, i.e. the prior of the relative deviation to be predicted;

[0131] Normal: normal distribution;

[0132] : The standard deviation of the observed value x, that is, the prior of the standard deviation to be predicted;

[0133] The conditional distribution of rd, that is, the posterior distribution of rd, is shown as follows:

[0134]

[0135] Where:

[0136] : The probability of relative deviation of the predicted concentration given X;

[0137] : is the total probability of rd;

[0138] : is the joint probability distribution of X and rd, that is, the marginal probability of rd under the condition of X;

[0139] According to the observed value x, the Bayesian method MCMC is used to calculate based on the pymc computing package , obtain the posterior probability distribution of the mean relative deviation of the predicted concentration, with its mean As the relative deviation of the concentration predicted by this method.

[0140] According to the above formula, the mean value of the relative deviation x of the high-precision compound is 0.68 (0.66-0.70), and the standard deviation is 0.16 (0.14-0.17). The relative deviation of 0.68 is taken as the predicted relative deviation of the high-precision compound.

[0141] The mean value of the relative deviation x of the low-precision compounds was 0.73 (0.64-0.81), and the standard deviation was 0.44 (0.37-0.50). The relative deviation of 0.73 was used as the predicted relative deviation of the high-precision compounds.

[0142] 6. Application

[0143] There are two application cases. The first case is a seawater sample. No blank sample was prepared, so the detection limit was not calculated. Among them, 8 compounds have both high-resolution mass spectrometry peak area information and liquid chromatography-triple quadrupole mass spectrometry quantitative results. Figure 4 As shown. In the figure, the red dots represent the quantitative results of liquid chromatography-triple quadrupole mass spectrometry, the blue dots represent the predicted concentrations, and the blue lines represent the upper and lower limits of the predictions. The red marks show the relative deviation of the predicted concentrations and the prediction accuracy. When the prediction accuracy is low, the prediction range is wider; when the prediction accuracy is high, the prediction range is narrower; the prediction ranges of compounds with high and low prediction accuracy both include the quantitative results of liquid chromatography-triple quadrupole mass spectrometry.

[0144] The second case is a sample from a sewage treatment plant in a certain industry. No blank sample was prepared, so the detection limit was not calculated. After pretreatment, the sample entered the high-resolution mass spectrometry non-targeted screening, and a total of 103 compounds were screened, including 64 high-precision compounds and 39 low-precision compounds. The predicted concentration results are shown in the following table. According to the subsequent requirements for concentration deviation, the predicted concentration, upper limit or lower limit can be selected for subsequent work.

[0145]

[0146]

[0147] In summary, the method of the present invention comprehensively considers the high-resolution mass spectrometry information, fingerprint information and chemical descriptors of the compound, and makes full use of neural networks and Bayesian statistical methods to effectively correct the errors caused by instrument response changes and pre-treatment processes, thereby maximizing the accuracy and reliability of concentration prediction. At the same time, an accuracy discrimination model is proposed, which provides a basis for subsequent optimization of the model.

[0148] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, the embodiments should be considered exemplary and non-restrictive in all respects, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims be included in the present invention.

[0149] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A method for non-targeted screening and identification of compound concentrations based on machine learning, characterized in that: The following steps are involved: S1. Constructing a trained deep residual neural network for predicting concentration, and obtaining the compound structure, mass spectrum fusion coding, predicted concentration and its predicted relative deviation during the training of the deep residual neural network; S2. constructing a trained compound prediction accuracy discrimination model and obtaining a reconstruction error of the compound prediction accuracy discrimination model during the training process; S3, obtaining structural information and mass spectrum information of the predicted compound in the spiked sample and the internal standard compound; S4, using the structural information and mass spectrum information of the predicted compound in the spiked sample and the internal standard compound as inputs to the trained deep residual neural network to obtain the compound structure, mass spectrum fusion code and predicted concentration of the predicted compound in the spiked sample and the internal standard compound; S5, outputting the predicted concentration of the trained deep residual neural network as the first concentration of the compound to be predicted; The compound structure and mass spectrum fusion coding of the spiked sample to be predicted compound and the internal standard compound are used as the input of the trained compound prediction accuracy discrimination model, and the accuracy of the concentration of the compound to be predicted is judged according to the reconstruction error of the compound prediction accuracy discrimination model during the training process; S6, obtaining an average relative response factor of the compound to be predicted relative to the internal standard compound according to the first concentration and peak area of ​​the compound to be predicted in the spiked sample and the first concentration and peak area of ​​the internal standard compound; S7, calculating the second concentration of the compound to be predicted based on the peak area and average relative response factor of the compound to be predicted in the spiked sample, the added amount and peak area of ​​the internal standard compound; S8, obtaining a relative deviation between the second concentration of the compound to be predicted and the relative spiked amount according to the added amount of the compound to be predicted and the second concentration of the compound to be predicted; S9. According to the accuracy information in S5, the relative deviation of the deep residual neural network model predicting a single compound during training is used as prior information for high-precision and low-precision compounds, and the relative deviation of the second concentration and the relative spiked amount of the compound to be predicted is used as posterior information. The lognormal distribution is used as the prior distribution and the normal distribution is used as the posterior distribution. The prediction relative deviations of high-precision and low-precision compounds are calculated by the Bayesian method respectively; S10, correcting the second concentration of the compound to be predicted by predicting the relative deviation to obtain the final concentration and range of the compound to be predicted; In S1, the method for constructing a deep residual neural network is specifically as follows: Compound features and peak areas are used as input, and concentration is used as output; a residual neural network model is built based on pytorch; MACCS molecular fingerprint spectra and Estate1 molecular fingerprint spectra are used as discrete feature inputs, and molecular descriptors are used as continuous feature inputs. After passing through one linear layer, they are fused into at least 6 subsequent layers of neural networks in the model, and then the peak area is input. Finally, the compound's escape features are output after passing through at least 2 layers of neural networks, and finally, the compound's concentration is output after passing through at least 2 layers of neural networks.

2. The method for non-targeted screening and identifying compound concentrations based on machine learning according to claim 1, characterized in that: In S1, the relative deviation of the model prediction for individual compounds The calculation formula is: Where: is the true concentration; is the predicted concentration; is the number of predicted concentrations of the same compound.

3. The method for non-targeted screening and identifying compound concentrations based on machine learning according to claim 1, characterized in that: The specific method of S6 is: Obtain at least 10 peak area-concentration pairs of the compound to be predicted and the internal standard compound, perform Monte Carlo simulation at least 50,000 times, and obtain the relative response factors of at least 5 peak area-concentration pairs each time RF Calculate and obtain the average relative response factor of the predicted compound relative to the internal standard compound ; Where: is the relative response factor, dimensionless; : The first concentration of the internal standard compound, μg / L; : The first concentration of the compound to be predicted in the spiked sample, μg / L; : Peak area of ​​internal standard compound; : Peak area of ​​the compound to be predicted; : average relative response factor, dimensionless; n: number of series points, n≥5.

4. The method for non-targeted screening and identifying compound concentrations based on machine learning according to claim 3, characterized in that: The specific methods of S7 and S8 are as follows: The compound to be predicted is set to multiple concentration gradients and added to pure water or other environmental samples respectively, with multiple groups of parallels set for each concentration gradient; after pre-treatment, the peak area is obtained by high-resolution mass spectrometry, and the prediction calculation is performed according to the following formula to obtain the second concentration of the compound to be predicted and the relative deviation of the relative spiked amount ; Where: is the second concentration of the compound to be predicted, μg / L; A: Peak area of ​​the compound to be predicted in the sample; : Peak area of ​​internal standard compound in sample; : The amount of internal standard compound added to the sample; Where: is the spike concentration; is the predicted concentration; is the number of spiked samples with the same spike concentration.

5. The method for non-targeted screening and identification of compound concentrations based on machine learning according to claim 1, characterized in that: In S9, the prior distribution of the relative deviations of high-precision and low-precision compounds is as follows: Where: rd: relative deviation probability distribution of high-precision or low-precision compounds, i.e., the prior distribution of relative deviation; Log-Normal: log-normal distribution; : estimated value of the rd mean; : rd is the estimated value of the standard deviation; : represents the average relative deviation of the i-th compound; : the number of compounds; Let the error event introduced by considering the instrument response and the pre-processing process be X, and let the relative deviation x obtained by S6 be the observed value of X; : The mean of the observed values ​​x, i.e. the prior of the relative deviation to be predicted; Normal: normal distribution; : The standard deviation of the observed value x, that is, the prior of the standard deviation to be predicted; The conditional distribution of rd, that is, the posterior distribution of rd, is shown as follows: Where: : The probability of relative deviation of the predicted concentration given X; : is the total probability of rd, that is, the probability of the relative deviation in S1; : is the joint probability distribution of X and rd, that is, the marginal probability of rd under the condition of X; According to the observed value x, the Bayesian method MCMC is used to calculate based on the pymc computing package , obtain the posterior probability distribution of the mean relative deviation of the predicted concentration, with its mean as the relative deviation from the predicted concentration.

6. The method for non-targeted screening and identifying compound concentrations based on machine learning according to claim 5, characterized in that: In S10, the final concentration range of the compound to be predicted is obtained according to the following formula; Where: : is the first concentration of the compound to be predicted; : is the relative deviation of the predicted concentration.

Images