An ai-assisted metabolomics differential analysis method
By using data preprocessing and AI analysis models, the batch effect and instrument drift problems in metabolomics differential analysis were solved, resulting in more accurate analysis results and higher statistical efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHANGHAI BIOTREE
- Filing Date
- 2026-06-12
- Publication Date
- 2026-07-14
AI Technical Summary
Existing metabolomics differential analysis methods, when using AI, struggle to effectively remove the effects of non-biological factors such as batch effects and instrument drift, leading to a decrease in the accuracy of the analytical results.
Data preprocessing techniques, including missing value imputation, logarithmic transformation, overall linear model fitting, standardization, and hyperparameter correction, are employed. Combined with inverse Gaussian distribution and maximum likelihood estimation, a difference analysis is performed using an AI analysis model to generate a complete analysis report.
It improves the accuracy of analysis results, avoids false positive and false negative results, enhances the ability of batch effect correction models to fit the true data distribution, and improves the statistical efficiency and robustness of parameter estimation.
Smart Images
Figure CN122392647A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of metabolomics analysis technology, and in particular to an AI-assisted method for differential metabolomics analysis. Background Technology
[0002] Metabolomics is a field that uses high-throughput detection and systematic analysis of low-molecular-weight metabolites in biological samples to reflect the metabolic fingerprint of an organism under specific physiological or pathological conditions. In recent years, metabolomics has played an increasingly important role in elucidating disease mechanisms, discovering biomarkers, early disease diagnosis, drug target identification, and precision medicine.
[0003] In current technologies, the mainstream methods for differential metabolomics analysis are still multivariate statistical methods, such as principal component analysis (PCA). As an unsupervised learning method, PCA projects high-dimensional metabolomics data onto a few principal components through dimensionality reduction, which is used for data quality assessment, outlier detection, and preliminary observation of overall clustering patterns. Partial least squares discriminant analysis (PLSA) achieves sample classification and feature selection by establishing a linear regression model between the metabolite matrix and classification labels. However, PCA has limited ability to discover differentially expressed metabolites associated with specific groups, and PLS is essentially a linear model, which makes it difficult to capture the complex nonlinear interactions between metabolites.
[0004] With the expansion of metabolomics research, the increase in data complexity, and the rapid development of AI technology, AI can be used to assist in differential analysis of metabolomics. However, the detection values of low molecular weight metabolites in biological samples can vary due to different batches of testing. When performing differential analysis using AI, this difference can greatly affect the analysis results. Therefore, how to ensure the accuracy of the analysis results when performing differential analysis using AI is the fundamental problem that this invention aims to solve. Summary of the Invention
[0005] To ensure the accuracy of the analysis results when using AI for differential analysis, this application provides an AI-assisted metabolomics differential analysis method, employing the following technical solution: An AI-assisted metabolomics differential analysis method includes: Perform data preprocessing on the biological group to obtain preprocessed data; Recognize user input commands and obtain the corresponding biological group and output content category based on the user input commands; AI analysis models are used to perform differential analysis on preprocessed data corresponding to biological groups to obtain differential analysis results and output content categories. A complete analysis report is generated based on the difference analysis results and the categories of output content.
[0006] Optionally, the data preprocessing process includes: Impute missing values and perform logarithmic transformation on the biological group to obtain n samples; group the n samples into batches to obtain K batch groups; An n×p matrix Y is constructed based on the metabolites of each sample, where p represents the metabolites; an n×q matrix X is constructed based on the number of biological covariates. Based on matrix X and matrix Y, an overall linear model is fitted, and the mean, standard deviation and residual of each metabolite are obtained. Standardize based on the mean, standard deviation and residual of metabolites, and calculate the batch mean and batch variance of each batch group after standardization; Estimate the hyperparameters of each batch based on the batch mean and batch variance; The hyperparameters of each batch are calibrated to obtain preprocessed data.
[0007] Optionally, the process of fitting the overall linear model includes: The estimated coefficients are obtained through ordinary least squares fitting. and the measured value of the i-th metabolite in the j-th sample ; For n samples Calculate the mean to obtain the mean of the i-th metabolite. ; according to Estimated coefficients And matrix X obtains the residual of the i-th metabolite in the j-th sample; The standard deviation of the i-th metabolite is obtained based on the residual. .
[0008] Furthermore, the standardization process includes: pass Calculate the value of the i-th metabolite in the j-th sample after standardization; For the k-th group Calculate the mean to obtain the batch mean. ; Based on batch average Calculate batch variance .
[0009] Optionally, the estimation process of the hyperparameters includes: Based on batch average Calculate the mean of all metabolites ,according to Calculate all metabolites The sample variance is used to determine the additive effect hyperparameters. Based on batch variance The marginal likelihood of a single metabolite is constructed using an inverse Gaussian distribution. An objective function is defined based on the marginal likelihood of a single metabolite, and the multiplicative effect hyperparameters are obtained by maximizing the objective function.
[0010] Optionally, the calibration process includes: Based on the hyperparameters, calculation models for the posterior mean of the additive effect and the posterior mean of the multiplicative effect are established respectively. The calculation models are iterated to obtain the posterior mean of the additive effect and the posterior mean of the multiplicative effect. Based on the posterior mean of additive effects and the posterior mode of multiplicative effects Perform calibration to obtain ; based on Perform an inverse standardization process to obtain the corrected measurement value. The corrected measurement value As preprocessed data.
[0011] Optionally, the training process of the AI analysis model includes: Missing values were handled and the sample data was standardized to determine the training and test sets; Perform feature filtering to obtain a feature subset; Define the loss function, perform internal validation, and repeat training; The model evaluation process is carried out to obtain the AI analysis model.
[0012] Optionally, the model evaluation process includes: Establish an ROC curve based on the model output results, and define the AUC value based on the area under the curve. Obtain the mean and standard deviation of the AUC values during the R independent data partitioning process; Perform a permutation test and obtain the p-value based on the results of the permutation test; Obtain the AUC values for the training and test sets respectively, and find the closest fit between the two. The model quality score is obtained based on the mean, standard deviation, p-value, and fit proximity value of the AUC value, and automatic optimization is performed based on the model quality score.
[0013] Optionally, the ROC curve is plotted with the false positive rate on the x-axis and the true positive rate on the y-axis; wherein, the true positive rate represents the proportion of the actual positive class that is correctly predicted at the current time point, and the false positive rate represents the proportion of the actual negative class that is incorrectly predicted as positive at the current time point.
[0014] In summary, this application includes at least one of the following beneficial technical effects: This invention, through data preprocessing, separates and removes systematic variations (influenced by non-biological factors such as batch effects, instrument drift, and differences in sample collection time) from the data, highlighting the true inter-group differences and avoiding false positive or false negative results, thus ensuring the accuracy of analysis results when AI performs differential analysis. By using inverse Gaussian and maximum likelihood estimation hyperparameters, since a few metabolites may exhibit significant scale variations due to batch effects in actual metabolomics data, the inverse Gaussian prior can preserve such signals without over-compression, improving the fidelity of the correction. Maximum likelihood estimation automatically adjusts the shape parameter of the inverse Gaussian distribution to fit the overall distribution of the data by maximizing the joint likelihood of all observations, thereby more accurately reflecting the true variability of batch effects. Ultimately, this significantly enhances the ability of the batch effect correction model to fit the real data distribution, improves the statistical efficiency and robustness of parameter estimation, and improves the false positive control and accuracy of downstream differential metabolite screening. Attached Figure Description
[0015] Figure 1 This is a flowchart of the AI-assisted metabolomics differential analysis method in this invention. Detailed Implementation
[0016] The embodiments of this application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.
[0017] In the description of this specification, the references to "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with the described embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0018] Please see Figure 1 This application discloses an AI-assisted metabolomics differential analysis method, comprising: 01: Perform data preprocessing on the biological group to obtain preprocessed data; 02: Recognize user input commands and obtain the corresponding biological group and output content category based on the user input commands; 03: Use AI analysis models to perform differential analysis on the preprocessed data corresponding to the biological group, and obtain the differential analysis results and output content categories; 04: Generate a complete analysis report based on the difference analysis results and output content categories.
[0019] In this embodiment, the biological group refers to different biological or experimental condition groups. The difference analysis is performed on two biological groups, such as the disease group and the healthy control group, or the drug treatment group and the solvent control group.
[0020] In one example, the process of metabolomics differential analysis is shown: 01: Perform data preprocessing on the Ctrl group and the Treat group to obtain the preprocessed data for the Ctrl group and the Treat group; 02: Obtain the user input command "Help me analyze the metabolomics differences between the Ctrl group and the Treat group, generate a volcano plot and enrichment analysis", and identify the following: task type: differential_analysis (differential analysis) (92%), business domain: metabolomics (95%), output format: chart+report, biological group: ["Ctrl group", "Treat group"], comparison dimension: group_comparison; 03: Perform differential analysis on the preprocessed data corresponding to the "Ctrl group" and "Treat group" using an AI analysis model to obtain the differential analysis results (a list of differential metabolites). Generate a visual volcano plot based on the differential analysis results, and perform KEGG / GO enrichment analysis based on the visual volcano plot.
[0021] 04: Generate a complete analysis report based on the difference analysis results, the visualized volcano plot, and the enrichment analysis results.
[0022] In one embodiment, a data preprocessing procedure is provided, including: 011: Perform missing value imputation and logarithmic transformation on the biological group to obtain n samples, indexed by j; group the n samples by batch to obtain K batch groups, indexed by k; where missing values are imputed with the median of the metabolite in the group; since raw metabolomics data are usually right-skewed (large kurtosis, variance increases with the mean), logarithmic transformation can make it closer to a normal distribution, satisfying the assumptions of subsequent parameter statistics, while stabilizing the variance; 012: Construct an n×p matrix Y based on the metabolites of each sample, where p represents the metabolites; construct an n×q matrix X, where q represents the number of biological covariates, such as age, sex, BMI, etc. 013: Based on matrices X and Y, a linear model is fitted to obtain the mean, standard deviation, and residuals for each metabolite. This process involves establishing a linear regression model for the i-th metabolite using its measurement vector across all samples as the dependent variable and matrix X as the independent variable. Fitting by ordinary least squares; Let be a column vector, representing the observed values of the i-th metabolite across all samples; For distribution error; obtain estimated coefficients and the measured value of the i-th metabolite in the j-th sample For n samples Calculate the mean to obtain the mean of the i-th metabolite. , ;according to Estimated coefficients And matrix X obtains the residual of the i-th metabolite in the j-th sample. , The standard deviation of the i-th metabolite is obtained based on the residual. , .
[0023] 014: Standardization is performed based on the mean, standard deviation, and residuals of metabolites. The standardization process includes: through... Calculate the value of the i-th metabolite in the j-th sample after standardization; for the k-th group... Calculate the mean to obtain the batch mean. Based on batch average Calculate batch variance , , For the set of sample indices within batch k, This represents the number of samples within batch k.
[0024] 015: Estimate the hyperparameters of each batch based on the batch mean and batch variance; the hyperparameter estimation process includes: estimating the hyperparameters based on the batch mean... Calculate the mean of all metabolites , ;according to Calculate all metabolites The sample variance is used to determine the additive effect hyperparameters. ; Based on batch variance And constructing the marginal likelihood of a single metabolite using an inverse Gaussian distribution: Where IG is an inverse Gaussian distribution, and x represents , Let x be the probability density function of the inverse Gaussian distribution; The objective function is defined based on the marginal likelihood of a single metabolite. ; The hyperparameters of the multiplicative effect are obtained by maximizing the objective function. In this embodiment, L-BFGS is used to maximize logL, Gauss-Legendary quadrature is used to calculate the integral of each i, and then the optimizer is called to solve for the result.
[0025] 016: Correction is performed based on the hyperparameters of each batch. The correction process includes: establishing calculation models for the additive effect posterior mean and the multiplicative effect posterior mean based on the hyperparameters. The calculation model for the additive effect posterior mean is as follows: ; The calculation model for the posterior mean of the multiplicative effect is as follows: The computational model is iterated, initially with indices i and k, let... Calculation obtained ; pass Recalculate the sum of squared residuals ,Will replace calculate Thus, the posterior mean of the additive effect is obtained. and the posterior mean of the multiplicative effect .
[0026] Based on the posterior mean of additive effects and the posterior mode of multiplicative effects Perform calibration to obtain , ;based on Perform an inverse standardization process to obtain the corrected measurement value. , ; the corrected measurement value As preprocessed data.
[0027] This embodiment of the application, through data preprocessing, can separate and remove systematic variations (interference from non-biological factors such as batch effects, instrument drift, and differences in sample collection time) from the data, highlighting the true inter-group differences and avoiding false positive or false negative results. This embodiment also uses inverse Gaussian and maximum likelihood estimation hyperparameters. Since in actual metabolomics data, a few metabolites may exhibit significant scale variations due to batch effects, the inverse Gaussian prior can preserve such signals without over-compression, improving the fidelity of the correction. Maximum likelihood estimation automatically adjusts the shape parameters of the inverse Gaussian distribution to fit the overall distribution of the data by maximizing the joint likelihood of all observations, thereby more accurately reflecting the true variability of batch effects. The table below compares the preprocessing method in this embodiment with the evaluation metrics of the standard ComBat: The comparison table of evaluation indicators shows that the smaller the value of the indicator corresponding to correction fidelity, the closer the correction effect is to the true baseline; the smaller the value of the indicator corresponding to effectiveness, the more thorough the batch effect removal; and the more sensitive the indicator corresponding to robustness, the more prone it is to bias. The method in this embodiment can better protect the biological differences in the data. The lower the value of the indicator corresponding to downstream analysis, the more statistically significant and reliable the results can be obtained when the data is used for downstream analysis. Therefore, the method in this embodiment can significantly enhance the ability of the batch effect correction model to fit the true data distribution, preserve the main biological signals more completely, improve the statistical efficiency and robustness of parameter estimation, and ultimately improve the false positive control and accuracy of downstream differential metabolite screening.
[0028] In one embodiment, a process for training an AI analysis model is provided, including: Missing values were handled and the sample data was standardized to determine the training and test sets; Perform feature filtering to obtain a feature subset; feature filtering can be achieved using methods such as filtering and embedding. Define the loss function, perform internal validation and repeated training; different AI base models can be selected according to the sample size, such as random forest (RF), support vector machine (SVM) and deep neural network (DNN); A model evaluation process is conducted to obtain an AI analysis model. In this embodiment, different AI models are trained and evaluated, and the final model is selected as the AI analysis model. The model evaluation process includes: ROC curves are constructed based on the model output results. The ROC curves are plotted with the false positive rate on the x-axis and the true positive rate on the y-axis. The true positive rate represents the proportion of the actual positive class that is correctly predicted at the current time point, and the false positive rate represents the proportion of the actual negative class that is incorrectly predicted as positive at the current time point. The AUC value is defined based on the area under the curve. Obtain the mean AUC value during R independent data partitioning processes. and standard deviation ; A permutation test is performed, and the p-value is obtained based on the test results. The permutation test process generates a null distribution by shuffling the labels, performing B permutations in total. Each permutation is indexed by b. After each random permutation, the same training and evaluation process is repeated to obtain the AUC value of the random permutation. ;pass Get the p value; Let be a conditional function, if If the value is 1, then the corresponding value is 1; otherwise, it is 0.
[0029] Obtain the AUC values for the training and test sets respectively. , , The AUC is the value obtained by evaluating the model on the training set itself after training it with m samples. The AUC is obtained by evaluating the model on an independent test set after training it with m samples, thus obtaining the closest fit between the two. ; The model quality score S is obtained based on the mean, standard deviation, p-value, and fit proximity value of the AUC. ; Automatic tuning and model selection are performed based on model quality scores. The model quality scores are compared with the quality threshold to quickly screen out reliable models for subsequent biological interpretation. The model with the highest quality score is selected to avoid the one-sidedness that may be caused by relying on a single indicator. At the same time, during the training process, the model quality score can be used as an objective function for automatic adjustment to improve the accuracy of the model.
[0030] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.
Claims
1. An AI-assisted method for differential metabolomics analysis, characterized in that, include: Perform data preprocessing on the biological group to obtain preprocessed data; Recognize user input commands and obtain the corresponding biological group and output content category based on the user input commands; AI analysis models are used to perform differential analysis on preprocessed data corresponding to biological groups to obtain differential analysis results and output content categories. A complete analysis report is generated based on the difference analysis results and the categories of output content.
2. The method for differential metabolomics analysis based on AI according to claim 1, characterized in that, The data preprocessing process includes: Impute missing values and perform logarithmic transformation on the biological group to obtain n samples; group the n samples into batches to obtain K batch groups; An n×p matrix Y is constructed based on the metabolites of each sample, where p represents the metabolites; an n×q matrix X is constructed based on the number of biological covariates. Based on matrix X and matrix Y, an overall linear model is fitted, and the mean, standard deviation and residual of each metabolite are obtained. Standardize based on the mean, standard deviation and residual of metabolites, and calculate the batch mean and batch variance of each batch group after standardization; Estimate the hyperparameters of each batch based on the batch mean and batch variance; The hyperparameters of each batch are calibrated to obtain preprocessed data.
3. The method for differential metabolomics analysis based on AI according to claim 2, characterized in that, The process of fitting the overall linear model includes: The estimated coefficients are obtained through ordinary least squares fitting. and the measured value of the i-th metabolite in the j-th sample ; For n samples Calculate the mean to obtain the mean of the i-th metabolite. ; according to Estimated coefficients And matrix X obtains the residual of the i-th metabolite in the j-th sample; The standard deviation of the i-th metabolite is obtained based on the residual. .
4. The method for differential metabolomics analysis based on AI according to claim 3, characterized in that, The standardization process includes: pass Calculate the value of the i-th metabolite in the j-th sample after standardization; For the k-th group Calculate the mean to obtain the batch mean. ; Based on batch average Calculate batch variance .
5. The method for differential metabolomics analysis based on AI according to claim 4, characterized in that, The estimation process of the hyperparameters includes: Based on batch average Calculate the mean of all metabolites ,according to Calculate all metabolites The sample variance is used to determine the additive effect hyperparameters. Based on batch variance The marginal likelihood of a single metabolite is constructed using an inverse Gaussian distribution. An objective function is defined based on the marginal likelihood of a single metabolite, and the multiplicative effect hyperparameters are obtained by maximizing the objective function.
6. The method for differential metabolomics analysis based on AI according to claim 5, characterized in that, The calibration process includes: Based on the hyperparameters, calculation models for the posterior mean of the additive effect and the posterior mean of the multiplicative effect are established respectively. The calculation models are iterated to obtain the posterior mean of the additive effect and the posterior mean of the multiplicative effect. Based on the posterior mean of additive effects and the posterior mode of multiplicative effects Perform calibration to obtain ; based on Perform an inverse standardization process to obtain the corrected measurement value. The corrected measurement value As preprocessed data.
7. The method for differential metabolomics analysis based on AI according to claim 1, characterized in that, The training process of the AI analysis model includes: Missing values were handled and the sample data was standardized to determine the training and test sets; Perform feature filtering to obtain a feature subset; Define the loss function, perform internal validation, and repeat training; The model evaluation process is carried out to obtain the AI analysis model.
8. The method for differential metabolomics analysis based on AI according to claim 7, characterized in that, The model evaluation process includes: Establish an ROC curve based on the model output results, and define the AUC value based on the area under the curve. Obtain the mean and standard deviation of the AUC values during the R independent data partitioning process; Perform a permutation test and obtain the p-value based on the results of the permutation test; Obtain the AUC values for the training and test sets respectively, and find the closest fit between the two. The model quality score is obtained based on the mean, standard deviation, p-value, and fit proximity value of the AUC value, and automatic optimization is performed based on the model quality score.
9. The method for differential metabolomics analysis based on AI according to claim 8, characterized in that, The ROC curve is plotted with the false positive rate on the x-axis and the true positive rate on the y-axis. The true positive rate represents the proportion of the actual positive class that is correctly predicted at the current time point, and the false positive rate represents the proportion of the actual negative class that is incorrectly predicted as positive at the current time point.