Machine learning-based unifi compound identification result classification and model construction method

Abstract

This invention discloses a machine learning-based method for classifying and building models of UNIFI compound identification results, belonging to the field of natural product cheminformatics. The method includes: acquiring a compound identification result table output by the UNIFI platform; manually verifying candidate compounds in the table and labeling them with binary classification supervision tags to form a training dataset; extracting a set of nine predetermined numerical features from the information of each candidate compound; using these nine features as input and the corresponding labels as targets, training a machine learning algorithm to obtain a classification model. In application, extracting the same nine features from a new identification result table and inputting them into the model will output the identification confidence category and matching score. This invention, through a fixed label and lightweight feature strategy, achieves rapid, objective, and automated classification of UNIFI high-throughput identification results, significantly improving verification efficiency and result consistency.

Description

Technical Field

[0001] This invention belongs to the field of natural product cheminformatics, specifically involving a method for classifying and constructing UNIFI compound identification results based on machine learning. Background Technology

[0002] Traditional Chinese medicines and health products are mostly extracted from plants, resulting in complex and diverse compositions. Various analytical methods are now available. Traditional analytical methods, such as spectroscopic and chromatographic methods, as well as the widely used modern high-performance liquid chromatography-mass spectrometry (HPLC-MS), can simultaneously provide retention time and mass-to-charge ratio information, offering advantages such as high sensitivity and efficiency. However, the manual analysis and annotation of massive amounts of data remains time-consuming and labor-intensive, and it is easy to miss valuable information.

[0003] The Waters UNIFI scientific information system integrates high-resolution mass spectrometry, unified software, and multi-source databases to provide a one-stop solution for compound identification, from data acquisition to results reporting. This system can automatically process MS data. E Mass spectrometry data, extracting and assigning parent and fragment ions, reduces the traditionally weeks-long manual spectral interpretation time to hours, significantly improving analytical throughput. However, this integrated algorithm still has inherent limitations. For example, it struggles to distinguish isomers with similar retention times or fragment spectra; low-content components may be filtered out due to response thresholds or background suppression; in complex fragmentation pathways, default fragment annotation rules may not match experimental spectra; furthermore, key parameters such as mass error and isotope tolerance require repeated adjustments based on human experience, which can easily lead to overly lenient or overly strict identification results. These shortcomings mean that after initial screening by UNIFI, researchers still need to spend a significant amount of time comparing each sample with standards or literature and performing highly subjective manual verification, making the entire process cumbersome and inefficient.

[0004] Therefore, the industry needs a solution that can maintain UNIFI's high-throughput advantage while effectively overcoming its discrimination shortcomings, and achieve fast, objective, and accurate automatic classification. Summary of the Invention

[0005] The technical problem to be solved by this invention is to overcome the limitations of the UNIFI system algorithm in complex sample analysis scenarios such as natural products, and to achieve rapid, objective, automated and accurate classification of compound identification results to replace inefficient and subjective manual verification.

[0006] To address the aforementioned issues, this invention proposes a machine learning-based method for classifying and building models for UNIFI compound identification results. This method utilizes a machine learning optimization strategy based on fixed labels and lightweight features. It directly employs "true / false" binary classification labels established through manual review as supervisory signals. Nine key numerical features are extracted from UNIFI identification results: mass error, mass spectrometry signal response, theoretical fragment hit count, secondary spectrum matching degree, peak area percentage, peak height percentage, chromatographic peak width ratio, root mean square ratio of isotope abundance, and root mean square error of isotope mass. A machine learning classification model is then used to achieve rapid model training on a standard computer, with training time controlled within one minute. The trained model can achieve one-click import and second-level classification of new batches of samples, automatically outputting reliable compound labels that are highly consistent with human judgment. While maintaining the high-throughput processing advantages of UNIFI, this method effectively improves the accuracy, objectivity, and efficiency of its identification results.

[0007] In a first aspect, the present invention provides a method for constructing a UNIFI compound identification result classification model based on machine learning, comprising the following steps:

[0008] Obtain the compound identification result table output by the UNIFI data processing platform, which contains information on multiple candidate compounds;

[0009] The candidate compounds in the identification result table are manually reviewed, and each candidate compound is labeled with a binary classification supervision label representing the identification confidence level to form a training dataset;

[0010] From the information of each candidate compound in the training dataset, a set of predetermined numerical features are extracted as model input. The set of predetermined numerical features consists of the following nine items: mass error, mass spectrometry signal response value, theoretical fragment ion hit count, secondary mass spectrum matching degree, chromatographic peak area percentage, chromatographic peak height percentage, chromatographic peak width ratio, isotope abundance matching root mean square ratio, and isotope mass matching root mean square error.

[0011] The extracted nine numerical features were used as model inputs, and the corresponding binary classification supervision labels were used as model training targets to train the machine learning algorithm and obtain a classification model for compound identification results.

[0012] Optionally, the manual verification is performed based on the following quantitative judgment criteria: the absolute value of the mass error is less than 10 ppm, the mass spectrometry signal response value is greater than 500, the theoretical fragment ion hit number is greater than 1, the matching degree of the secondary mass spectrum is greater than 0%, the percentage of chromatographic peak area is greater than 0%, the percentage of chromatographic peak height is greater than 0%, the chromatographic peak width ratio is less than or equal to 2.0, the root mean square ratio of isotope abundance matching is less than or equal to 8, and the root mean square error of isotope mass matching is less than 10 ppm.

[0013] Optionally, when training the machine learning algorithm, the training process is balanced according to the proportion of samples of different confidence categories in the training data.

[0014] Optionally, the machine learning algorithm includes a gradient boosting decision tree model.

[0015] Secondly, the present invention provides a method for classifying UNIFI compound identification results, comprising:

[0016] The compound identification result classification model is constructed by applying the machine learning-based UNIFI compound identification result classification model construction method described above;

[0017] For the UNIFI compound identification results table to be classified, nine numerical features are extracted for each candidate compound; the nine numerical features include: mass error, mass spectrometry signal response value, theoretical fragment ion hit number, secondary mass spectrum matching degree, chromatographic peak area percentage, chromatographic peak height percentage, chromatographic peak width ratio, isotope abundance matching root mean square ratio, and isotope mass matching root mean square error.

[0018] The extracted features are input into the compound identification result classification model, which outputs the identification confidence category for each candidate compound.

[0019] Optionally, the classification method for the UNIFI compound identification results further includes: calculating and outputting a matching score for each candidate compound based on the classification probability output by the compound identification result classification model.

[0020] Thirdly, the present invention provides an automatic classification system for compound identification results, comprising:

[0021] The data interface module is used to obtain the compound identification result table output by the UNIFI data processing platform;

[0022] The label management module is used to support manual verification and binary labeling of candidate compounds in the identification result table to form a training dataset;

[0023] The feature extraction module is used to extract a set of predetermined numerical features from the candidate compound information in the training dataset. The set of predetermined numerical features consists of the following nine items: mass error, mass spectrometry signal response value, theoretical fragment ion hit count, secondary mass spectrum matching degree, chromatographic peak area percentage, chromatographic peak height percentage, chromatographic peak width ratio, isotope abundance matching root mean square ratio, and isotope mass matching root mean square error.

[0024] The model building module is used to train a machine learning algorithm based on nine numerical features and corresponding binary classification supervised labels to obtain a classification model for compound identification results.

[0025] The classification application module is used to load the compound identification result classification model, automatically extract nine numerical features from the new UNIFI compound identification result table and classify them, and output the identification confidence category.

[0026] Fourthly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, is configured to perform the following steps:

[0027] Obtain the compound identification result table output by the UNIFI data processing platform, and manually verify and label the candidate compounds in the table to form a training dataset;

[0028] From the candidate compound information in the training dataset, a set of predetermined numerical features is extracted. The set of predetermined numerical features consists of the following nine items: mass error, mass spectrometry signal response value, theoretical fragment ion hit count, secondary mass spectrum matching degree, chromatographic peak area percentage, chromatographic peak height percentage, chromatographic peak width ratio, isotope abundance matching root mean square ratio, and isotope mass matching root mean square error.

[0029] A machine learning algorithm is trained based on nine numerical features and their corresponding labels to obtain a classification model for compound identification results; and / or, the compound identification result classification model is loaded, nine numerical features are extracted from the new UNIFI identification result table and input into the model, and the identification confidence category is output.

[0030] Fifthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor and configured to perform the following steps:

[0031] Obtain the compound identification result table output by the UNIFI data processing platform, and manually verify and label the candidate compounds in the table to form a training dataset;

[0032] From the candidate compound information in the training dataset, a set of predetermined numerical features is extracted. The set of predetermined numerical features consists of the following nine items: mass error, mass spectrometry signal response value, theoretical fragment ion hit count, secondary mass spectrum matching degree, chromatographic peak area percentage, chromatographic peak height percentage, chromatographic peak width ratio, isotope abundance matching root mean square ratio, and isotope mass matching root mean square error.

[0033] A machine learning algorithm is trained based on nine numerical features and their corresponding labels to obtain a classification model for compound identification results; and / or, the compound identification result classification model is loaded, nine numerical features are extracted from the new UNIFI identification result table and input into the model, and the identification confidence category is output.

[0034] The beneficial effects of this invention are:

[0035] 1. To address the issues of subjective evaluation results and inconsistent standards, this invention employs manual review based on fixed quantitative standards and binary label annotation. This approach transforms expert experience into reproducible objective supervisory signals, ensuring the consistency and reliability of labels from the data source, and providing a high-quality, low-bias supervisory data foundation for subsequent model training.

[0036] 2. To address the limited discrimination capability of the native UNIFI algorithm in complex samples, this invention creatively defines and extracts a core feature set consisting of nine predetermined numerical features, including mass error, mass spectrometry signal response value, and theoretical fragment ion hit count. These features are directly related to key quality dimensions for compound identification, eliminating redundant information and complex transformations. This allows the machine learning model to accurately focus on the core indicators affecting the reliability of identification, thereby significantly improving the model's ability to distinguish between genuine and counterfeit compounds.

[0037] 3. To address the issues of low efficiency and difficulty in batch processing of manual review, this invention trains a lightweight machine learning model based on the aforementioned features and labels, and deploys an automated classification process for new data, achieving a fundamental shift from manual, item-by-item review to model-based batch judgment within seconds. In practice, classifying identification result tables containing hundreds of compounds can be completed in just a few seconds, while traditional manual review typically takes hours or even longer. This solution not only significantly improves analytical throughput but also enables efficient training and rapid application on ordinary computing devices, significantly reducing the threshold and time cost of review work.

[0038] In summary, this invention, through a collaborative technical system of "fixed standard annotation," "selected core features," and "lightweight model training and automated application," not only significantly improves the objectivity, accuracy, and consistency of compound identification result verification, but also realizes a paradigm shift from manual operation to intelligent execution, providing a time-saving, reliable, and efficient automated solution for high-throughput mass spectrometry data analysis. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 The overall flowchart of the machine learning classification and model construction method for UNIFI compound identification results provided by this invention.

[0041] Figure 2 This is a schematic diagram of the classification performance evaluation results obtained from the model hyperparameter selection and optimization process provided by the present invention.

[0042] Figure 3 This diagram illustrates the classification performance evaluation results obtained by cross-validation on the training set for the model provided in this invention. Detailed Implementation

[0043] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0044] Example 1

[0045] This embodiment provides a machine learning classification and model building method for UNIFI compound identification results. The core of this method lies in directly utilizing the structured identification result table output by the UNIFI platform, based on manually labeled fixed binary classification labels and a set of predefined key numerical features, to train a classification model through machine learning that can automatically and quickly distinguish the reliability of identification results, and achieve automated batch classification of new data.

[0046] like Figure 1 As shown, firstly, ultra-high performance liquid chromatography-quadrupole time-of-flight mass spectrometry (UPLC-QTOF-MS) was used to acquire LC-MS data of the samples. As an example, chromatographic separation was performed using an ACQUITY UPLC BEH C18 column with gradient elution using an aqueous solution containing 0.1% formic acid and acetonitrile as the mobile phase. Mass spectrometry acquisition was performed in negative ion mode with MSE data-dependent acquisition, and the mass scan range was set to m / z 50 to 1500.

[0047] Subsequently, qualitative analysis of the compounds in the sample LC-MS data was performed using the Waters UNIFI scientific information processing platform. The qualitative analysis criteria were as follows: background was automatically subtracted from the sample data, and peak width and peak area thresholds were automatically detected; the high-energy channel intensity threshold was set to 50 counts, and the secondary mass spectrometry intensity threshold was set to 200 counts; the mass tolerance was set to ±5 mDa, the fragment ion mass tolerance was ±10 mDa, the fragment structure prediction score was allowed to be lower than 8, all fragment ion information was retained, the extracted ion chromatogram mass tolerance was ±5 mDa, only the best-matching compounds were displayed, and a maximum of two compounds with the same mass-to-charge ratio were matched; for positive ion mode adducts, [M+H]+, [M+Na]+, and [M+NH4]+ were selected, and for negative ion mode adducts, [MH]- and [M+HCOO]- were selected; the correction mass-to-charge number was set to m / z 556.2766 for positive ion mode and m / z 554.2620 for negative ion mode.

[0048] After processing, the UNIFI platform outputs an identification results table containing detailed information on all candidate compounds, which can be exported as a single Excel file. The results table includes the following information: Row No., Item Name, Component Name, Formula, Identification Status, Item Tag, Observed m / z, Neutral mass (Da), Observed neutral mass (Da), MS Response, Mass error (mDa), Mass error (ppm), Adducts, Mass peak resolution, Theoretical Fragments Found, Total Fragments Found, Percentage of Product Spectrum Accounted For (%), Observed RT (min), Retention time (min), Area, Height, and Total Area. area, total height, % area (%), % height (%), chromatographic width (min), chromatographic width ratio, threshold parameter, inflection width (s), peak width parameter (s), 2nd derivative apex (min), root mean square percentage of isotope match intensity (RMS Percent), root mean square error of isotope match mass-to-charge ratio (RMS PPM).

[0049] Next, experienced analysts manually review the preliminary UNIFI identification results. This review employs a pre-defined set of fixed numerical judgment criteria to ensure the objectivity and consistency of the labeling. Specifically, the judgment is primarily based on the following nine numerical indicators that can be directly obtained from the UNIFI results table, and each indicator has fixed threshold requirements:

[0050] Quality error (usually expressed in ppm): The value must be less than 10, and the smaller the absolute value, the more accurate the quality measurement.

[0051] Mass spectrometry signal response value: The value must be greater than 500, and the higher the response value, the better the signal quality is generally.

[0052] Theoretical fragment ion hit count: The value must be greater than 1, and the more theoretical fragment ions that match, the stronger the identification evidence. A value of 10 indicates the strongest identification evidence.

[0053] Secondary mass spectrum matching degree: The value must be greater than 0, and the higher the percentage, the better the agreement between the experimental spectrum and the theoretical spectrum. 100 is the best.

[0054] Chromatographic peak area percentage and peak height percentage: These reflect the relative content and peak shape of the component in the chromatogram. The area percentage value should be greater than 0, and the larger the better, with 100 being the best. The height percentage value should also be greater than 0, and the larger the better, with 100 being the best.

[0055] Peak width ratio: The ratio must be less than or equal to 2.0, and the closer the ratio is to 1, the more symmetrical the peak shape and the more ideal the chromatographic behavior. In this invention, the peak width ratio refers to a parameter directly calculated and output by UNIFI software during peak detection. Specifically, it is the ratio of the actual full width of the chromatographic peak of the component to the average full width of the full width of the retention time of all chromatographic peaks in the same batch of analysis. Using this parameter can effectively identify and exclude components with abnormal conditions such as peak tailing, splitting, or co-elution.

[0056] Isotope distribution matching indices include the root mean square ratio of isotope abundance (the root mean square value of the ratio of measured to theoretical isotope abundance, reflecting the accuracy of the isotope response) and the root mean square error of isotope mass matching (unit: ppm, the root mean square value of the deviation between measured and theoretical isotope mass, reflecting the accuracy of isotope mass quality). Both are used together to evaluate the degree of agreement between measured and theoretical isotope models. Specifically, the root mean square percentage of isotope abundance should be less than or equal to 8, and the lower the better; the root mean square error of isotope mass should be less than 10 ppm, and the lower the better.

[0057] During manual evaluation, a comprehensive assessment is conducted based on fixed thresholds for the aforementioned nine indicators. To ensure the reliability of the identification results, this review standard requires that all nine indicators of the candidate compound simultaneously meet their respective threshold requirements. If any indicator exceeds the threshold, the compound is deemed "unreliable." The order of the nine features reflects their importance in the reliability assessment, but the final review conclusion is based on the "AND" logic (simultaneous satisfaction) of all thresholds. The review process follows a procedure where one researcher conducts initial screening and labeling according to the above standards, followed by independent, item-by-item review by at least another researcher. When review opinions differ, discussions or a third expert's decision are held until a consensus is reached. This mechanism aims to minimize individual subjective bias and ensure the objectivity and reliability of the generated training data labels.

[0058] Analysts comprehensively considered the above indicators and combined them with their professional experience to classify each candidate compound into "credible" or "uncredible" categories. This classification result was then used as a supervised learning label, such as "true" for "credible" and "false" for "uncredible," and directly appended to the original identification result table, thus forming a labeled training dataset.

[0059] After obtaining the labeled training dataset, feature engineering is performed. This invention abandons the strategy of using all available fields or complex derived features, creatively selecting only nine general, quantifiable core numerical features as input to the machine learning model. These nine features are: mass error (unit: ppm), mass spectrometry signal response value, theoretical fragment ion hit count, secondary mass spectrum matching degree (unit: %), chromatographic peak area percentage (unit: %), chromatographic peak height percentage (unit: %), chromatographic peak width ratio, isotope abundance matching root mean square ratio, and isotope mass matching root mean square error (unit: ppm).

[0060] The training dataset was read using programming tools. Nine feature values ​​were extracted for each compound record and matched with its manually labeled value to form the final dataset required for model training. Since the number of "unreliable" compounds in real-world data often far exceeds the number of "reliable" compounds, a class imbalance problem exists. During model training, a class weighting strategy was adopted, such as setting class weights based on the inverse ratio of the number of positive and negative class samples in the training data, thereby balancing the model's attention to different confidence levels.

[0061] In this embodiment, the nine extracted feature values ​​are all numerical data that can be directly used for model training. Their magnitude differences are within the tolerance range of the selected machine learning algorithm; therefore, no additional data cleaning, normalization, or standardization processing was performed. Simultaneously, the training dataset was manually reviewed to ensure completeness and the absence of missing feature values. Those skilled in the art should understand that, depending on the specific algorithm or data, conventional data preprocessing steps can also be employed without departing from the core concept of this invention.

[0062] Subsequently, this embodiment utilizes machine learning algorithms for model training. These algorithms can be gradient boosting decision tree models such as LightGBM or XGBoost. This example preferably uses the LightGBM algorithm, which has advantages in handling numerical features, dealing with imbalanced data, and achieving rapid training. The aforementioned nine-dimensional feature vector is used as the model input, and the corresponding binary classification labels are used as the training target. The learning task is set as binary classification, and the area under the curve (AUC) is selected as the evaluation metric.

[0063] To optimize model performance, different combinations of hyperparameters, including learning rate, maximum tree depth (max_depth), and number of leaves (num_leaves), were examined. Specifically, the learning rate was set to 0.05, 0.1, and 0.74, and the combinations of (max_depth, num_leaves) were set to (5, 20), (6, 31), and (7, 50), respectively. Accuracy, recall, F1 score, and AUC were used for evaluation and comparison (see [link to relevant documentation]). Figure 2 , Figure 3 After comprehensive evaluation, the optimal parameter combination for the LightGBM model is determined as follows:

[0064] (1) objective: set to "binary", indicating a binary classification task;

[0065] (2) metric: set to “auc”, indicating that the evaluation metric is AUC;

[0066] (3) learning_rate: Set to 0.1, representing the learning rate;

[0067] (4) max_depth: Set to 6, representing the maximum depth of the tree;

[0068] (5) num_leaves: Set to 31, representing the number of leaves;

[0069] (6) feature_fraction: Set to 0.8, which means that 80% of the features are used in each iteration (column sampling);

[0070] (7) bagging_fraction: Set to 0.8, which means that 80% of the samples are used in each iteration (row sampling);

[0071] (8) bagging_freq: Set to 5, which means that sample sampling is performed once every 5 iterations;

[0072] (9) scale_pos_weight: Set the calculation method to (len(labels)-sum(labels)) / sum(labels), which means that the weight of positive samples is automatically set to balance the classes;

[0073] (10) seed: Set to 123 to set a random seed to ensure that the results are reproducible;

[0074] An early stopping strategy is adopted during training, which means that training is stopped if there is no significant improvement in the performance on the validation set for 50 consecutive rounds, in order to prevent overfitting.

[0075] In a specific application, a dataset containing 1077 labeled compound records was used, with 192 records labeled "true" and 885 records labeled "false". This dataset was divided into training and test sets at an 80% to 20% ratio, and a 5-fold cross-validation method was employed for model training and validation. The entire training process could be completed within one minute on a typical personal computer. Testing showed that the model exhibited excellent classification performance on the test set: the average accuracy, recall, and F1 score were all greater than 0.93, and the average AUC was higher than 0.99, indicating that the model has a high classification and discrimination ability for the UNIFI platform's identification results.

[0076] Once training is complete, a savable and reusable classification model for compound identification results can be obtained, which can be used to automatically classify the UNIFI identification results of new batches of samples.

[0077] During the application phase, the system automatically reads the newly exported identification results table (Excel format) from the UNIFI platform, extracts the same nine numerical features as those used in the training phase for each candidate compound, and inputs them into the loaded classification model. The model outputs the predicted probability of each compound belonging to the "credible" category. To further verify the model's generalization ability, 20 completely independent batches of liquid chromatography-mass spectrometry (LC-MS) data identification results (a total of 3128 compounds) were used as an external validation set for testing. See the table below:

[0078] Table 1. Example of partial data representation of the external validation set.

[0079]

[0080] It should be noted that Table 1 only lists a portion of the data as an example; the actual data scale is as described above (3128 compounds). Furthermore, the fields listed in Table 1 are examples of raw information output from the UNIFI platform or model input features. Specifically, the component number is the record number in the validation set; the sample name is the experimental batch identifier; the component name is the preliminary identification result from the UNIFI system; the identification status, molecular formula, adduct, and measurement m / z are all system output information; and the quality error (ppm) is one of the key input features used for model classification. This Table 1 aims to demonstrate the composition of the raw validation set data; the model's output results (i.e., prediction confidence category and matching score) are not shown here.

[0081] The model successfully screened out 138 high-confidence components (labeled as "true") from 3128 compounds. The matching score of these 138 compounds was higher than 60.00, while the scores of the remaining compounds were lower than 60.00, and the labels and scores matched consistently.

[0082] During the application phase, the system automatically reads the newly exported identification results table (Excel format) from the UNIFI platform, extracts the same nine numerical features as those used in the training phase for each candidate compound, and inputs them into the loaded classification model. The model outputs the predicted probability of each compound belonging to the "credible" category. To further verify the model's generalization ability and classification accuracy, the identification results of 20 completely independent batches of sample liquid chromatography-mass spectrometry data were used as an external validation set. This validation set contains 3128 compounds, which have been independently reviewed and labeled by professionals without referring to the model results. Among them, 148 compounds were labeled as "credible" (true), and 2980 compounds were labeled as "uncredible" (false), forming an objective evaluation benchmark.

[0083] Statistical analysis showed that the model's predictions on the validation set achieved an accuracy of 98.69% compared to the human review baseline, a recall rate (for truly reliable compounds) of 88.41%, and a precision of 92.99%. Specifically, the model automatically selected 138 high-confidence components (prediction probability > 0.6, matching score > 60.00) from 3128 compounds, of which 122 matched human judgment; the remaining 2990 were judged as low-confidence by the model (score ≤ 60.00), of which 2965 matched human judgment. This demonstrates that the method of this invention can efficiently and accurately reproduce expert review conclusions.

[0084] The system makes classification decisions based on predicted probabilities and generates a final result. Typically, compounds with predicted probabilities higher than a default decision threshold are classified as "trustworthy," otherwise as "untrustworthy." In this embodiment, a decision threshold of 0.5 is preferred. Simultaneously, the probability value of a compound being predicted as "trustworthy" (denoted as ) can be... Perform linear mapping The results are converted into a matching score from 0 to 100, which intuitively reflects the reliability of the identification results. The entire process is fully automated and requires no human intervention. Practice has shown that for a new identification result table containing hundreds of compounds, the total time from importing the Excel file to outputting the complete classification results and scoring report is only within a few seconds, achieving truly one-click rapid verification.

[0085] Example 2

[0086] This embodiment provides an automatic classification system for compound identification results that can implement the above method. The system may include the following functional modules:

[0087] The data interface module is used to obtain the compound identification result table output by the UNIFI data processing platform;

[0088] The label management module is used to support manual verification and binary labeling of candidate compounds in the identification result table to form a training dataset;

[0089] The feature extraction module is used to extract a set of predetermined numerical features from the candidate compound information in the training dataset. The set of predetermined numerical features consists of the following nine items: mass error, mass spectrometry signal response value, theoretical fragment ion hit count, secondary mass spectrum matching degree, chromatographic peak area percentage, chromatographic peak height percentage, chromatographic peak width ratio, isotope abundance matching root mean square ratio, and isotope mass matching root mean square error.

[0090] The model building module is used to train a machine learning algorithm based on nine numerical features and corresponding binary classification supervised labels to obtain a classification model for compound identification results.

[0091] The classification application module is used to load the compound identification result classification model, automatically extract nine numerical features from the new UNIFI compound identification result table and classify them, and output the identification confidence category.

[0092] Example 3

[0093] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it is configured to perform the following steps:

[0094] Obtain the compound identification result table output by the UNIFI data processing platform, and manually verify and label the candidate compounds in the table to form a training dataset;

[0095] From the candidate compound information in the training dataset, a set of predetermined numerical features is extracted. The set of predetermined numerical features consists of the following nine items: mass error, mass spectrometry signal response value, theoretical fragment ion hit number, secondary mass spectrum matching degree, chromatographic peak area percentage, chromatographic peak height percentage, chromatographic peak width ratio, isotope abundance matching root mean square ratio, and isotope mass matching root mean square error.

[0096] A machine learning algorithm is trained based on nine numerical features and their corresponding labels to obtain a classification model for compound identification results; and / or, the classification model for compound identification results is loaded, nine numerical features are extracted from the new UNIFI identification result table and input into the model, and the identification confidence category is output.

[0097] Example 4

[0098] This embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is configured to perform the following steps:

[0099] Obtain the compound identification result table output by the UNIFI data processing platform, and manually verify and label the candidate compounds in the table to form a training dataset;

[0100] From the candidate compound information in the training dataset, a set of predetermined numerical features is extracted. The set of predetermined numerical features consists of the following nine items: mass error, mass spectrometry signal response value, theoretical fragment ion hit number, secondary mass spectrum matching degree, chromatographic peak area percentage, chromatographic peak height percentage, chromatographic peak width ratio, isotope abundance matching root mean square ratio, and isotope mass matching root mean square error.

[0101] A machine learning algorithm is trained based on nine numerical features and their corresponding labels to obtain a classification model for compound identification results; and / or, the classification model for compound identification results is loaded, nine numerical features are extracted from the new UNIFI identification result table and input into the model, and the identification confidence category is output.

[0102] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for constructing a UNIFI compound identification result classification model based on machine learning, characterized in that, Includes the following steps: Obtain the compound identification result table output by the UNIFI data processing platform, which contains information on multiple candidate compounds; The candidate compounds in the identification result table are manually reviewed, and each candidate compound is labeled with a binary classification supervision label representing the identification confidence level to form a training dataset; From the information of each candidate compound in the training dataset, a set of predetermined numerical features are extracted as model input. The set of predetermined numerical features consists of the following nine items: mass error, mass spectrometry signal response value, theoretical fragment ion hit count, secondary mass spectrum matching degree, chromatographic peak area percentage, chromatographic peak height percentage, chromatographic peak width ratio, isotope abundance matching root mean square ratio, and isotope mass matching root mean square error. The nine extracted numerical features were used as model inputs, and the corresponding binary classification supervision labels were used as model training targets to train the machine learning algorithm and obtain a classification model for compound identification results. The machine learning algorithm used is the LightGBM algorithm. The learning task is set as binary classification, and the area under the curve is used as the evaluation metric. Different combinations of hyperparameters, including learning rate, maximum tree depth and number of leaves, are examined, and accuracy, recall, F1 score and area under the curve are used for evaluation and comparison.

2. The method of claim 1, wherein the machine learning-based UNIFI compound identification result classification model is constructed by using a machine learning algorithm. The manual verification is based on the following quantitative judgment criteria: the absolute value of the mass error is less than 10 ppm, the mass spectrometry signal response value is greater than 500, the theoretical fragment ion hit number is greater than 1, the matching degree of the secondary mass spectrum is greater than 0%, the percentage of chromatographic peak area is greater than 0%, the percentage of chromatographic peak height is greater than 0%, the chromatographic peak width ratio is less than or equal to 2.0, the root mean square ratio of isotope abundance matching is less than or equal to 8, and the root mean square error of isotope mass matching is less than 10 ppm. 3.The method of claim 1, wherein the UNIFI compound identification result classification model is constructed based on machine learning. When training the machine learning algorithm, the training process is balanced according to the proportion of samples of different confidence categories in the training data.

4. The method of claim 1, wherein the machine learning-based UNIFI compound identification result classification model is constructed. The machine learning algorithm includes a gradient boosting decision tree model.

5. A method of classifying UNIFI compound identification results, comprising: include: The compound identification result classification model is constructed using the method for constructing the UNIFI compound identification result classification model based on machine learning as described in any one of claims 1 to 4; For the UNIFI compound identification results table to be classified, nine numerical features are extracted for each candidate compound; The nine numerical characteristics include: mass error, mass spectrometry signal response value, theoretical fragment ion hit count, secondary mass spectrum matching degree, chromatographic peak area percentage, chromatographic peak height percentage, chromatographic peak width ratio, root mean square ratio of isotope abundance matching, and root mean square error of isotope mass matching. The extracted features are input into the compound identification result classification model, which outputs the identification confidence category for each candidate compound.

6. The classification method for UNIFI compound identification results according to claim 5, characterized in that, The classification method for the UNIFI compound identification results further includes: calculating and outputting a matching score for each candidate compound based on the classification probability output by the compound identification result classification model.

7. A system for automatically classifying compound identification results, comprising: include: The data interface module is used to obtain the compound identification result table output by the UNIFI data processing platform; The label management module is used to support manual verification and binary labeling of candidate compounds in the identification result table to form a training dataset; The feature extraction module is used to extract a set of predetermined numerical features from the candidate compound information in the training dataset. The set of predetermined numerical features consists of the following nine items: mass error, mass spectrometry signal response value, theoretical fragment ion hit count, secondary mass spectrum matching degree, chromatographic peak area percentage, chromatographic peak height percentage, chromatographic peak width ratio, isotope abundance matching root mean square ratio, and isotope mass matching root mean square error. The model building module is used to train a machine learning algorithm based on nine numerical features and corresponding binary classification supervised labels to obtain a classification model for compound identification results. The classification application module is used to load the compound identification result classification model, automatically extract nine numerical features from the new UNIFI compound identification result table and classify them, and output the identification confidence category. The machine learning algorithm used is the LightGBM algorithm. The learning task is set as binary classification, and the area under the curve is used as the evaluation metric. Different combinations of hyperparameters, including learning rate, maximum tree depth and number of leaves, are examined, and accuracy, recall, F1 score and area under the curve are used for evaluation and comparison.

8. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the computer program, it is configured to perform the following steps: Obtain the compound identification result table output by the UNIFI data processing platform, and manually verify and label the candidate compounds in the table to form a training dataset; From the candidate compound information in the training dataset, a set of predetermined numerical features is extracted. The set of predetermined numerical features consists of the following nine items: mass error, mass spectrometry signal response value, theoretical fragment ion hit count, secondary mass spectrum matching degree, chromatographic peak area percentage, chromatographic peak height percentage, chromatographic peak width ratio, isotope abundance matching root mean square ratio, and isotope mass matching root mean square error. A machine learning algorithm is trained based on nine numerical features and their corresponding labels to obtain a compound identification result classification model; and / or, the compound identification result classification model is loaded, nine numerical features are extracted from the new UNIFI identification result table and input into the model, and the identification confidence category is output. The machine learning algorithm used is the LightGBM algorithm. The learning task is set as binary classification, and the area under the curve is used as the evaluation metric. Different combinations of hyperparameters, including learning rate, maximum tree depth and number of leaves, are examined, and accuracy, recall, F1 score and area under the curve are used for evaluation and comparison.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, When the computer program is executed by the processor, it is configured to perform the following steps: Obtain the compound identification result table output by the UNIFI data processing platform, and manually verify and label the candidate compounds in the table to form a training dataset; From the candidate compound information in the training dataset, a set of predetermined numerical features is extracted. The set of predetermined numerical features consists of the following nine items: mass error, mass spectrometry signal response value, theoretical fragment ion hit count, secondary mass spectrum matching degree, chromatographic peak area percentage, chromatographic peak height percentage, chromatographic peak width ratio, isotope abundance matching root mean square ratio, and isotope mass matching root mean square error. A machine learning algorithm is trained based on nine numerical features and their corresponding labels to obtain a compound identification result classification model; and / or, the compound identification result classification model is loaded, nine numerical features are extracted from the new UNIFI identification result table and input into the model, and the identification confidence category is output. The machine learning algorithm used is the LightGBM algorithm. The learning task is set as binary classification, and the area under the curve is used as the evaluation metric. Different combinations of hyperparameters, including learning rate, maximum tree depth and number of leaves, are examined, and accuracy, recall, F1 score and area under the curve are used for evaluation and comparison.

Images