A machine learning based method and system for non-target screening of organosilicon contaminants

Abstract

The application discloses a kind of based on machine learning's organic silicon pollutant non-target screening method and system, it is related to environmental non-target substance screening technical field, comprising: obtaining the mass spectrum data of target to be measured;Characteristic engineering processing is carried out to mass spectrum data and obtains first data set;Ring line classification is carried out to target to be measured, and the classification result is coupled first data set and obtains second data set;Molecular formula data of target to be measured is obtained, and coupled second data set and obtains third data set;Random forest model is sequentially trained based on first data set, second data set and third data set respectively;Random forest model is adjusted to super parameter;Output result.The application is suitable for siloxane and other new pollutants, applies machine learning to the mass spectrum comparison method based on prior data, is used for screening and identification of silicon-containing organic matter in complex environmental medium, reduces the difficulty of method based on mass spectrum information spectrum analysis, and organic silicon pollutant is screened quickly and efficiently.

Description

Technical Field

[0001] This invention relates to the field of environmental non-target substance screening technology, and more specifically to a machine learning-based method and system for non-target screening of organosilicon pollutants. Background Technology

[0002] In recent years, organosilicon polymers have exhibited excellent properties in terms of temperature resistance, weather resistance, electrical properties, radiation resistance, surface properties, repairability, and safety and environmental friendliness (low flammability, low toxicity, odorless, physiological inertness, and human-friendly properties), leading to their large-scale synthesis. However, during the use and production of organosilicon polymer products, residual small-molecule monomers may enter the environment through various pathways. Therefore, with the large-scale production and use of organosilicon, attention has been paid to the toxic and environmental effects of monomers such as siloxanes. For example, methylsiloxanes D4 and D5 are difficult to degrade in animals and exhibit accumulation characteristics. Furthermore, studies have shown that methylsiloxanes have reproductive, immunological, and neurotoxic effects on animals.

[0003] Currently, there are numerous types of organosilicones. Taking siloxanes as an example, in addition to the common dimethylsiloxane, special branched siloxanes such as phenyl, trifluoropropyl, amino, and vinyl siloxanes are also widely produced and emitted into the environment. However, due to limitations in analytical methods, it is impossible to conduct a comprehensive assessment of the full composition of organosilicon compounds in the environment. In order to comprehensively and reasonably evaluate the environmental effects of organosilicones, non-target screening has gradually developed and is used as the main means of screening unknown substances in complex environmental matrices. At present, high-resolution mass spectrometry such as GC-QTOF has the potential to discover non-target pollutants due to its high-resolution full-scan mode and accurate mass numbers, providing a theoretical possibility for high-throughput full-component analysis of organosilicones. However, traditional pollutant screening methods are based on spectral comparison, determination of the molecular formula and structural formula of substances, and finally verification with standards. This method is not suitable for new pollutants such as siloxanes. Currently, there is a lack of research on its properties and toxicity. Mass spectrometry libraries, such as the NIST mass spectrometry library, only have spectral analyses of a little over a hundred siloxanes, which is a fatal flaw for screening organosilicones. At the same time, this method of analyzing spectroscopic information based on mass spectrometry is time-consuming and labor-intensive, making it difficult to screen organosilicon contaminants quickly and efficiently. For some media containing numerous organosilicon samples, if the spectroscopic analysis is performed on each one, the screening difficulty and workload are enormous.

[0004] Mass spectrometry, a method based on prior data, is perfectly suited for machine learning and has been successfully applied in proteomics and metabolomics. However, to date, no one has used machine learning for non-target screening of organosilicon utilization. Therefore, proposing a machine learning-based non-target screening method for organosilicon pollutants to screen and identify silicon-containing organic compounds in complex environmental media, and achieving rapid and efficient screening, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides a non-target screening method and system for organosilicon pollutants based on machine learning, for screening and identifying silicon-containing organic compounds in complex environmental media, solving the problem of the difficulty in spectroscopic interpretation of siloxanes. To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A machine learning-based method for non-target screening of organosilicon contaminants, comprising:

[0007] Acquire mass spectrometry data of the target object;

[0008] The first dataset is obtained by performing feature engineering on the mass spectrometry data;

[0009] A random forest model is constructed. The random forest model is trained on a first dataset to obtain the cyclic classification results of the target to be tested. The cyclic classification results are coupled with the first dataset to obtain a second dataset. The random forest model is trained on the second dataset to obtain the molecular formula data of the target to be tested. The molecular formula data are coupled with the second dataset to obtain a third dataset. The random forest model is trained on the third dataset to obtain the group category of the target to be tested.

[0010] Hyperparameter tuning for the random forest model;

[0011] Output results.

[0012] Optionally, the feature engineering includes precise mass difference and peak intensity screening of silicon isotopes.

[0013] Optionally, the specific steps of the ring-line classification are as follows: convert the ring-line into a binary representation, with the ring marked as 0 and the line marked as 1.

[0014] Optionally, the specific steps for obtaining the molecular formula data of the target to be tested are as follows: obtaining the molecular formula data of the target to be tested, and decomposing the molecular formula of the target to be tested into the number of atoms, including the number of atoms of C, H, O and Si.

[0015] Optionally, the third dataset includes the first dataset, the ring-linear classification results of the target to be tested, and the molecular formula data of the target to be tested. The model's label is a group classification label, including pure methyl without special groups, amino, phenyl, vinyl, fluorine-containing, hydroxyl, and H atoms directly bonded to Si, which are represented by numbers from 0 to 6.

[0016] Optionally, the random forest model is built using the sklearn third-party library in the Python language environment, with 80% of the dataset used for training and 20% for testing.

[0017] Optionally, the specific steps for training the random forest model are as follows: the random forest model is trained using a first dataset and a classification model is selected for ring-linear classification prediction; it is trained using a second dataset and a regression model is selected to predict the number of atoms; and it is trained using a third dataset and a multi-classification model is selected to predict the group category.

[0018] Optionally, the specific steps for hyperparameter tuning of the random forest model are as follows: using the GridSearchCV interface in the sklearn library to find the optimal parameters through grid search, training the learner with the tuned parameters, and finding the parameters with the highest accuracy on the test set from all the parameters.

[0019] Optionally, the output results may also include a comparison and verification between traditional non-target screening results and machine learning results to obtain the training effect of the random forest model.

[0020] Optionally, a computer-readable storage medium stores a computer program that, when executed by a processor, implements a machine learning-based non-target screening method for organosilicon contaminants.

[0021] Optionally, a machine learning-based non-target screening system for organosilicon contaminants includes:

[0022] Acquisition module: Used to acquire mass spectrometry data of the target object;

[0023] Processing module: Constructs a random forest model; trains the random forest model on a first dataset to obtain the cyclic classification result of the target object, and couples the cyclic classification result with the first dataset to obtain the second dataset; trains the random forest model on the second dataset to obtain the molecular formula data of the target object, and couples the molecular formula data with the second dataset to obtain the third dataset; trains the random forest model on the third dataset to obtain the group category of the target object.

[0024] Tuning module: Used to tune the hyperparameters of the random forest model;

[0025] Output module: Used to output results.

[0026] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a non-target screening method and system for organosilicon pollutants based on machine learning, which has the following beneficial effects:

[0027] This invention switches the ring-linear classification to binary representation to facilitate model learning;

[0028] This invention is applicable to novel pollutants such as siloxanes. It applies machine learning to a mass spectrometry comparison method based on prior data for the screening and identification of silicon-containing organic pollutants in complex environmental media. The mass spectrometry data undergoes feature engineering to obtain a first dataset, which is then used for ring-shaped classification prediction. The classification results are coupled with the first dataset to obtain a second dataset, which is used for molecular formula prediction of the target. The molecular formula prediction results are then coupled with the second dataset to obtain a third dataset, which is used for group category prediction. A random forest model is sequentially trained based on the first, second, and third datasets to obtain the optimal random forest model. This reduces the spectral complexity of mass spectrometry-based methods, enabling rapid and efficient screening of organosilicon pollutants.

[0029] This invention mimics the hierarchical relationship of a neural network through a multi-layer machine learning model. First, mass spectrometry data is acquired. The supplementary content of the second dataset is the output result of the first-layer model. The supplementary content of the third dataset is also the output result of the second-layer model. By feeding back the model results sequentially, the dataset size is expanded, improving the prediction effect on the number of atoms and the classification of functional groups, and increasing the accuracy. Attached Figure Description

[0030] 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.

[0031] Figure 1 This is a schematic diagram of the structure provided by the present invention.

[0032] Figure 2 This is a schematic diagram of the process provided by the present invention. Detailed Implementation

[0033] 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.

[0034] This invention discloses a machine learning-based method for screening non-target organosilicon contaminants, comprising:

[0035] Acquire mass spectrometry data of the target object;

[0036] The first dataset is obtained by performing feature engineering on the mass spectrometry data;

[0037] A random forest model is constructed. The random forest model is trained on a first dataset to obtain the cyclic classification results of the target to be tested. The cyclic classification results are coupled with the first dataset to obtain a second dataset. The random forest model is trained on the second dataset to obtain the molecular formula data of the target to be tested. The molecular formula data are coupled with the second dataset to obtain a third dataset. The random forest model is trained on the third dataset to obtain the group category of the target to be tested.

[0038] Hyperparameter tuning for the random forest model;

[0039] Output results.

[0040] Furthermore, the specific steps for training the random forest model are as follows: the random forest model is trained using a first dataset and a classification model is selected for ring-linear classification prediction; it is trained using a second dataset and a regression model is selected to predict the number of atoms; and it is trained using a third dataset and a multi-classification model is selected to predict the group category.

[0041] Furthermore, feature engineering is performed on the mass spectrometry data to obtain the first dataset, which is used for predicting the ring-shaped classification results. The classification results are coupled with the first dataset to obtain the second dataset, which is used for predicting the molecular formula of the target. The molecular formula prediction results are coupled with the second dataset to obtain the third dataset, which is used for predicting the group category. The random forest model is then trained sequentially based on the first, second, and third datasets. Further, the optimal parameters are found using the GridSearchCV interface in the sklearn library, and the learner is trained using the adjusted parameters, identifying the parameters with the highest accuracy on the test set from all available parameters.

[0042] Furthermore, the output results also include a comparison and verification between traditional non-target screening results and machine learning results to determine the training effect of the random forest model.

[0043] In a specific implementation, a machine learning-based non-target screening method for organosilicon contaminants, such as... Figure 1 As shown:

[0044] S1: Purchase and configure the standard, and perform sample analysis on a GC-Orbitrap-MS instrument (Trace1310-Q ExactiveGC, Thermo Fisher Scientific, USA). The instrument settings are as follows: TG-5 SILMS column (30m×0.25mm I.D×0.25μm Film, MaxTemp: 320 / 350℃), helium constant flow rate of 1.2mL / min, pulse splitless mode at 280℃, injection volume of 1μL, GC column oven initial temperature of 40℃ and held for 2 minutes, increased to 100℃ at 20℃ / min, then increased to 200℃ at 10℃ / min and held for 5 minutes, and finally increased to 280℃ at 20℃ / min and held for 3 minutes. Mass spectrometry is performed in high resolution mode, ion source and transfer line temperature are both 280℃, solvent delay is 3 minutes, resolution is 6000, AGCtarget1e6, and mass scan range is 40-600 amu.

[0045] S2: Samples are analyzed using Orbitrap, peak intensities and relative mass-to-nucleus ratios are obtained, and a mass spectrum peak list is exported. Based on the precise mass differences of isotopes 0.9996 (29Si-28Si), 0.9973 (30Si-29Si), and 1.9968 (30Si-28Si) amu (Audiand Wapstra, 1995), a Python script is used to filter a large amount of mass spectrometry data. Due to instrument error, a range of 0.99954-0.99960 is set, and peaks with mass differences within this range are considered to contain silicon.

[0046] S3: To prevent too many features from affecting the model, the relative mass-to-nucleus ratio and peak intensity after screening are exported. Python is used to traverse the list and screen out the m / z peak intensities that are greater than a specific value. For each standard, 9-10 peaks with the highest m / z intensity are screened. Missing values ​​are filled with 0. Then, the molecular mass of the substance is obtained using a CI source. Finally, the features are integrated to form the feature part of the machine learning dataset.

[0047] S4: For siloxane compounds, the general structures are divided into cyclic and linear, with atoms including C, H, O, and Si. Modifying groups include trifluoropropyl, vinyl, phenyl, etc. Therefore, the dependent variables for the label settings are as follows: cyclic and linear (binary conversion), number of C, number of H, number of O, number of Si, and classification of groups such as methyl, trifluoropropyl, vinyl, phenyl, and amino (numerical classification).

[0048] S5: A random forest model is used. This model can handle high-dimensional, large-feature data, and because the feature subset is randomly selected, feature selection is unnecessary. After processing the data, it can identify which features are important. The model has strong generalization ability, fast data processing speed, and is easily parallelized (trees are independent during training), making it relatively simple to implement. Different models are selected for different labels; for example, classification models are used for toroidal classification and group classification, while a regression model is used to predict the number of atoms.

[0049] In Example 1, machine learning combined with non-target screening was used to address the lack of mass spectra of organosilicon substances, achieving rapid and efficient screening. After training all models, actual environmental samples were tested. Surface water samples (0.5-1.0 m deep) were collected using 1L glass bottles pre-washed with n-hexane. A 4cm*2cm polyethylene film was added to the glass bottle as an adsorbent to reduce siloxane volatilization, and the samples were stored in a headspace-free sealed container. Bottom sediment samples (0-15 cm deep) were collected using a stainless steel grab-type sediment sampler pre-washed with n-hexane. Each sample was a mixture of five randomly sampled samples and stored in a sealed polyethylene bag. All surface water and bottom sediment samples were stored at 4°C and pretreated within 48 hours.

[0050] Sample pretreatment: Take 100 mL of filtered water sample, add 40 μL of internal standard solution (M4Q) dissolved in acetone, extract once with 25 mL of n-hexane solution, centrifuge, and transfer the upper organic layer to a flat-bottomed flask. Then, perform a second extraction with 20 mL of a mixed solvent of n-hexane / ethyl acetate (volume ratio 1:1). Subsequently, the organic layer is transferred to a flask and concentrated to 8-10 mL under nitrogen. The extract is purified by passing it through a 1.0 g anhydrous sodium sulfate column to remove residual water, then purged to 0.5 mL under nitrogen, and brought to a final volume of 1 mL with n-hexane / ethyl acetate (volume ratio 1:1). Finally, the extract is transferred to a vial for analysis.

[0051] The mass spectrometry data of the sample were fed into the model for calculation. The model achieved an overall accuracy of 70% with the previous data. In the actual sample testing, M4Q was used as an unknown substance for testing. Only the number of H atoms differed by 1. The other classification and atom number regression were accurate. Therefore, it is confirmed that machine learning is feasible in the field of non-target screening of organosilicon and is suitable for rapid and efficient assessment of pollutants in environmental media.

[0052] Furthermore, a machine learning-based non-target screening system for organosilicon contaminants includes:

[0053] Acquisition module: Used to acquire mass spectrometry data of the target object;

[0054] Processing module: Constructs a random forest model; trains the random forest model on a first dataset to obtain the cyclic classification result of the target object, and couples the cyclic classification result with the first dataset to obtain the second dataset; trains the random forest model on the second dataset to obtain the molecular formula data of the target object, and couples the molecular formula data with the second dataset to obtain the third dataset; trains the random forest model on the third dataset to obtain the group category of the target object.

[0055] Tuning module: Used to tune the hyperparameters of the random forest model;

[0056] Output module: Used to output results.

[0057] 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.

[0058] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. 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 non-target screening method for organosilicon contaminants based on machine learning, characterized in that, include: Acquire mass spectrometry data of the target object; The first dataset is obtained by performing feature engineering on the mass spectrometry data; The feature engineering includes precise mass difference and peak intensity screening of silicon isotopes; A random forest model is constructed. The random forest model is trained on a first dataset to obtain the cyclic classification results of the target to be tested. The cyclic classification results are coupled with the first dataset to obtain a second dataset. The random forest model is trained on the second dataset to obtain the molecular formula data of the target to be tested. The molecular formula data are coupled with the second dataset to obtain a third dataset. The random forest model is trained on the third dataset to obtain the group category of the target to be tested. The specific steps for training the random forest model are as follows: the random forest model is trained using the first dataset, and a classification model is selected for cyclic linear classification prediction. The number of atoms was predicted by training on the second dataset and using a regression model. The group category was predicted by training on a third dataset and using a multi-classification model. Hyperparameter tuning for the random forest model; Output results.

2. The non-target screening method for organosilicon contaminants based on machine learning according to claim 1, characterized in that, The specific steps for the ring-line classification are as follows: transform the ring-line into a binary representation, with the ring marked as 0 and the line marked as 1.

3. The non-target screening method for organosilicon contaminants based on machine learning according to claim 1, characterized in that, The specific steps for obtaining the molecular formula data of the target to be tested are as follows: obtain the molecular formula data of the target to be tested, and decompose the molecular formula of the target to be tested into the number of atoms, including the number of atoms of C, H, O and Si.

4. The non-target screening method for organosilicon contaminants based on machine learning according to claim 1, characterized in that, The third dataset includes the first dataset, the ring-linear classification results of the target to be tested, and the molecular formula data of the target to be tested. The model's label is a group classification label, including pure methyl without special groups, amino, phenyl, vinyl, fluorine-containing, hydroxyl, and H atoms directly bonded to Si, which are represented by numbers from 0 to 6.

5. The non-target screening method for organosilicon contaminants based on machine learning according to claim 1, characterized in that, The random forest model is built using the sklearn third-party library in the Python language environment, with 80% of the dataset used for training and 20% used for testing.

6. The non-target screening method for organosilicon contaminants based on machine learning according to claim 1, characterized in that, The specific steps for hyperparameter tuning of the random forest model are as follows: using the GridSearchCV interface in the sklearn library to find the optimal parameters through grid search, training the learner with the tuned parameters, and finding the parameters with the highest accuracy on the test set from all the parameters.

7. The non-target screening method for organosilicon contaminants based on machine learning according to claim 1, characterized in that, The output also includes a comparison and verification between traditional non-target screening results and machine learning results to determine the training effect of the random forest model.

8. A machine learning-based non-target screening system for organosilicon contaminants, characterized in that, include: Acquisition module: Used to acquire mass spectrometry data of the target object; The first dataset is obtained by performing feature engineering on the mass spectrometry data; The feature engineering includes precise mass difference and peak intensity screening of silicon isotopes; Processing module: Constructing a random forest model; The random forest model is trained on the first dataset to obtain the loop-shaped classification result of the target to be tested. The loop-shaped classification result is coupled with the first dataset to obtain the second dataset. The random forest model is trained on the second dataset to obtain the molecular formula data of the target to be tested, and the molecular formula data is coupled with the second dataset to obtain the third dataset; the random forest model is trained on the third dataset to obtain the group category of the target to be tested; The specific steps for training the random forest model are as follows: the random forest model is trained using a first dataset, and a classification model is selected for ring-linear classification prediction; the model is then trained using a second dataset, and a regression model is selected to predict the number of atoms. The group category was predicted by training on a third dataset and using a multi-classification model. Tuning module: Used to tune the hyperparameters of the random forest model; Output module: Used to output results.

Images