Peripheral blood metabonomics-based lymphoma early diagnosis model constructed by using neural network and application

Abstract

The invention discloses a lymphoma early diagnosis marker based on metabonomics and artificial intelligence technology and application thereof. A serum metabonomics technology and an artificial intelligence data analysis technology are applied, a Grad-CAM model is used for visualization, and the biomarker suitable for early lymphoma auxiliary diagnosis and an early lymphoma auxiliary diagnosis model are obtained. The model construction method is simple and convenient, and the obtained diagnosis model is excellent in effect, high in sensitivity, excellent in specificity and suitable for early auxiliary diagnosis of lymphoma. Diagnosis can be completed only through blood sampling detection, additional tissue sample collection is not needed, the diagnosis process is simple and rapid, early discovery and timely treatment of lymphoma are facilitated, and the kit has high clinical application and popularization value.

Description

Technical Field

[0001] The present invention relates to the field of medical data processing and deep learning, and specifically to the construction of an early diagnosis model for lymphoma based on peripheral blood metabolomics using a neural network and its application. Background Art

[0002] Lymphoma is a cancer that originates in the lymphatic system. Its epidemiological characteristics are diverse and vary significantly between different populations and different geographical regions (Smith et al., 2020). The clinical features of lymphoma vary from asymptomatic lymphadenopathy to systemic symptoms such as fever (Johnson & Federico, 2019). Given the heterogeneity of lymphoma, early and accurate diagnosis is crucial for the effective management of patients and the effectiveness of treatment.

[0003] Metabolomics, the comprehensive analysis of small molecule metabolites in biological systems, has become a powerful tool in the field of oncology (Lee et al., 2018). It provides insights into metabolic alterations that accompany cancer progression and provides potential biomarkers for disease diagnosis and prognosis (Harris et al., 2017).

[0004] Traditional metabolomics analysis faces various challenges, such as dealing with complex high-dimensional data, requiring extensive expertise, and computational limitations of traditional methods. Advanced computational methods, such as machine learning, are currently being developed to address these challenges (Kourou et al., 2015; Zhang et al., 2019). However, metabolomics data are one-dimensional sequences that detail the concentration of specific components, lack the temporal properties of signals such as electrocardiograms or electroencephalograms, and cannot be converted into interpretable two-dimensional images by methods such as Fourier or wavelet transforms. Therefore, specialized analysis methods are necessary. In order to address the challenges of applying traditional deep learning visualization models to metabolomics, the present invention discloses a method for metabolomics data classification, model construction, and feature visualization using a one-dimensional neural network, aiming to achieve early detection and identification of specific metabolic features that can serve as diagnostic markers. Summary of the invention

[0005] In view of the problems existing in the prior art, the purpose of the present invention is to provide a model and application for early diagnosis of lymphoma based on peripheral blood metabolomics using neural networks. The present invention uses artificial intelligence technology to construct a high-precision prediction model that can distinguish lymphoma patients from healthy controls and identify a group of specific metabolites associated with lymphoma. The model described in the present invention is not only statistically significant, but also provides a reasonable biological explanation through bioinformatics analysis, providing a new perspective for studying the pathogenic mechanism of lymphoma.

[0006] In order to achieve the above-mentioned object of the invention, the present invention provides the following technical solutions.

[0007] The present invention discloses an early auxiliary diagnosis marker for lymphoma, characterized in that the marker is the following (1) or / and (2): (1) a combination of 10 serum metabolites, monoacylglycerol, phosphatidylserine, androgens and their derivatives, oxidized phosphoglycerates, steroid derivatives, steroids, glycerophospholipids, oxidized phosphoglycerol phosphates, medium-chain hydroxy acids and derivatives, and phosphatidylserine; (2) A combination of 10 serum metabolites: phosphatidic acid, diglyceride, phosphatidic acid, decaglycerol, methionine, androgens, α-hydroxyprogesterone, testosterone acetate, dehydroepiandrosterone acetate, and methionine.

[0008] Furthermore, the above-mentioned lymphoma early auxiliary diagnosis marker is used in the preparation of lymphoma early auxiliary diagnosis products.

[0009] The present invention also discloses the use of a substance for detecting the above-mentioned lymphoma early auxiliary diagnosis marker in the preparation of a lymphoma early auxiliary diagnosis product.

[0010] Furthermore, the substance is a substance used to detect the content of diagnostic markers in serum.

[0011] The present invention also discloses a lymphoma early auxiliary diagnosis kit, characterized in that the kit contains a substance for detecting the lymphoma early auxiliary diagnosis marker described in claim 1.

[0012] Furthermore, the kit also includes a carrier recording the model or a device loaded with the model.

[0013] The present invention also discloses a method for constructing a lymphoma AI diagnostic model, which is characterized by comprising the following steps: (1) Serum samples were measured by UPLC-Q / TOF MS to obtain serum metabolites; (2) The metabolites obtained in (1) were normalized and preprocessed to have zero mean and unit variance; (3) A simplified ResNet network is used for one-dimensional convolution calculation, and the loss function uses cross entropy loss; (4) Use K-Fold cross validation to obtain the best parameters and the best model for feature visualization; (5) After obtaining feature activation data using Grad-CAM++, plot feature importance curves and use heat maps; (6) Arrange the feature weights calculated by Grad-CAM++ from large to small according to the numerical value to obtain the marker described in claim 1.

[0014] The present invention also discloses a lymphoma AI diagnostic model, characterized in that it is constructed by the lymphoma AI diagnostic model construction method according to claim 7.

[0015] The present invention also discloses a computing device, characterized in that it includes a memory and a processor, wherein the memory stores a disease screening model that can be loaded and executed by the processor to implement the method for constructing a lymphoma AI diagnostic model as described in claim 7.

[0016] The present invention also discloses a computer-readable storage medium, characterized in that it stores a disease screening model that can be loaded and executed by a processor to implement the method for constructing a lymphoma AI diagnostic model as described in claim 8.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows.

[0018] The present invention uses peripheral blood metabolomics technology and artificial intelligence data analysis technology to determine biomarkers and early lymphoma diagnostic models suitable for early lymphoma diagnosis. The biomarker screening method of the present invention is highly operational, the model construction method is simple, the obtained diagnostic model has excellent effect, high sensitivity, and excellent specificity, and is suitable for early diagnosis of lymphoma. The present invention can complete the diagnosis only by blood sampling and testing, without the need for additional tissue sample collection, which greatly reduces the risk of trauma. The diagnostic process of the present invention is simple and rapid, which is conducive to the early detection and timely treatment of lymphoma, and has high clinical application and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The overall structure of the present invention.

[0020] Figure 2 Schematic diagram of the K-Fold cross-validation principle.

[0021] Figure 3 Visualization curves and heatmaps of individual features of negative ion data.

[0022] Figure 4 Average feature visualization curve and heat map of negative ion data.

[0023] Figure 5 ROC curves of 10 validations of negative ion marker combinations. Figures A to J are validation curves for each validation.

[0024] Figure 6 ROC curves of 10 validations of the positive ion marker combination. Figures A to J are validation curves for each validation. DETAILED DESCRIPTION

[0025] The present invention is further described in detail below with specific embodiments, but it should not be understood that the scope of the above subject matter of the present invention is limited to the following embodiments, and all technologies realized based on the content of the present invention belong to the scope of the present invention.

[0026] Unless otherwise specified, all reagents and materials used in the present invention are commercially available.

[0027] Example 1: Detection of metabolites in lymphoma serum samples.

[0028] 1. Materials and reagents.

[0029] Acetonitrile (LC-MS grade), formic acid (LC-MS grade) and methanol (LC-MS grade) were purchased from Merck Chemical Technology (Shanghai) Co., Ltd. Leucine-enkephalin was purchased from Waters Technology (Shanghai) Co., Ltd.

[0030] 2. Instruments and equipment.

[0031] Waters ACQUITY UPLC TM Ultra-high performance liquid chromatograph (Waters Co., Milford, MA, USA), Xevo TM G2-XS QTof time-of-flight mass spectrometer (Waters MS Technologies, Manchester, UK), Masslynx TM NT 4.1 workstation; micropipettes (Eppendorf, Hamburg, Germany).

[0032] 3. Sample pretreatment.

[0033] 3.1 Serum samples.

[0034] The serum samples were thawed in a refrigerator (4°C) and then vortexed for 1 min. 120 μL of serum samples were added to 480 μL of methanol-acetonitrile solution (v / v=1:1, precooled to -20°C) to precipitate proteins, vortexed for 30 s, sonicated for 1 min, and then left at -20°C for 1 h to fully precipitate proteins. g Centrifuge for 10 min, collect the supernatant into a new 1.5 mL EP tube, and blow dry under nitrogen at room temperature. Resuspend the sample residue in 120 μL methanol-water solution (v / v=1:1, precooled to -20°C) and vortex for 30 s, and sonicate in an ice-water bath for 2 min. g After centrifugation for 10 min, the supernatant was filtered through a 0.22 μm filter membrane into a vial insert for UPLC-Q / TOF MS analysis.

[0035] 10 μL of all serum samples from each experimental group were mixed to obtain a mixed serum sample, which was pretreated according to the above method to obtain a serum quality control sample (QC).

[0036] 4. UPLC-Q / TOF MS measurement conditions.

[0037] Chromatographic separation was performed using Waters ACQUITY UPLC TM Ultra-high performance liquid chromatography system (Waters Corp, Milford, MA, USA). Chromatographic column: Waters ACQUITY HSS T3 column (100 mm × 2.1 mm id, 1.8 μm); column temperature: 40 °C; automatic injection sample chamber temperature: 4 °C; flow rate: 0.4 mL / min; injection volume: 2 μL in positive ion mode, 4 μL in negative ion mode. Mobile phase: binary gradient system, phase A: 0.1% formic acid-acetonitrile (v / v); phase B: 0.1% formic acid-water (v / v). The gradient elution program for plasma / serum sample detection was as follows: 0-1.0 min, 10-20% A; 1.0-3.5 min, 20-30% A; 3.5-4.0 min, 30-40% A; 4.0-6.0 min, 40-50% A; 6.0-6.5 min, 50-58% A; 6.5-11.0 min, 58-100% A; 11.0-14.0 min, 100 % A; 14.0-14.1 min, 100-10% A; 14.1-16.0 min, 10% A.

[0038] Mass spectrometry was performed using a XevoTM G2-XS QTof quadrupole time-of-flight mass spectrometer (Waters MS Technologies, Manchester, UK). MS conditions: electrospray ionization source (ESI); positive and negative ion modes; scan range: m / z 50-1000. In LockSpray mode, mass calibrant: leucine enkephalin (m / z 556.2771 [M+H]+, m / z 554.2615 [MH]-), concentration: 200 pg / μL, flow rate: 10 μL / min. Positive ion capillary voltage: 3.0 kV, negative ion capillary voltage: -2.5 kV; cone voltage: 40 V; source temperature: 120°C; desolvation temperature: 500°C; cone gas flow rate: 50 L / h; desolvation gas flow rate: 800 L / h. Data acquisition was performed using MasslynxTMNT4.1 software, data acquisition mode: Centroid; MS scanning methods are as follows: (1) MS EExperiment, channel 1: m / z 50-1000, 0.2 s scan time, 0.02 s interval scan delay, 6 V collision energy. Channel 2: m / z 50-1000, 0.2 s scan time, 0.02 s interval scan delay, collision energy range: 10-40 eV.

[0039] 5. Data preprocessing.

[0040] QI software (Waters, USA) was used to extract ion pairs, align peaks, match peaks, and calibrate peak intensity on the raw data (raw. file) spectra collected by UPLC-Q / TOF MS. The main parameters were set as follows: mass range 100-1000 amu, mass tolerance 0.02, minimum intensity 1%, mass window 0.02, retention time window 0.20, and noise elimination level 6. After applying 80% filtering rules to the obtained data table to remove zero values ​​and missing values, the data was normalized to the total ion intensity of each chromatogram and saved as new data. Finally, a three-dimensional matrix consisting of the specified peak sequence number (corresponding to the retention time and mass-to-charge ratio), observation point (sample number), and normalized peak intensity was obtained. This matrix will be used for all subsequent data analysis.

[0041] Embodiment 2: Model establishment.

[0042] The enhanced data set containing 120 samples is divided into a training set and a test set, where 100 samples in the entire data set are the training set and 20 samples are the test set.

[0043] First, the data is preprocessed and normalized; then the feature extraction and classification model is built, the neural network is pruned based on the ResNet network, and the loss function uses the cross entropy loss; cross validation is used to obtain the best parameters and train the best model; finally, the features extracted by the best model are used to construct the feature importance graph.

[0044] The specific steps are as follows.

[0045] 1. Get the data.

[0046] The data were processed using medical methods and dedicated software (as described in Example 1).

[0047] 120 negative ion data were generated, including 58 negative samples (non-disease) and 62 positive samples (patients). Each data contains 8786 features.

[0048] 125 positive ion data were generated, including 60 negative samples (non-disease) and 65 positive samples (patients). Each data contains 17179 features.

[0049] 2. Build a model to extract TOPN.

[0050] Step 1: Data preprocessing.

[0051] The metabolomics data of peripheral blood tests obtained from the hospital have a large range of numerical variations, which is not conducive to model convergence. Therefore, the input is Z-score standardized to have zero mean and unit variance.

[0052] Step 2: Build the model.

[0053] Because there are many data features and significant segments need to be extracted, a simplified ResNet network is used for one-dimensional convolution calculation to ensure classification performance while improving training efficiency. The network structure is shown in the table, where L represents the sequence length.

[0054] Table 1 Details of the modified convolutional neural network structure .

[0055] The loss function uses cross entropy loss, and the binary cross entropy loss formula is: in For sample The label of the positive sample is 1 and the negative sample is 0. For sample The probability of predicting a positive result.

[0056] Step 3: Perform cross-validation and train the best model.

[0057] Due to the limited number of samples, in order to determine the hyperparameters with the best training effect, the present invention uses the K-Fold cross-validation method. It divides the training set into K folds (parts), and then cycles K times, each time selecting one fold as validation data and the remaining folds as training data. This method can more accurately evaluate the performance of the model, avoid overfitting or underfitting problems, and also balance the effects of training and testing, providing more reliable model performance evaluation results. The schematic diagram of K-Fold cross-validation is shown in the figure. Figure 2 shown.

[0058] The present invention uses 5-fold cross validation. After training multiple times through this method, after determining the optimal hyperparameters such as learning rate, activation function, dropout, etc., this set of hyperparameters is adopted, and the model is trained for the last time using all the data in the training set to obtain the optimal model for feature visualization.

[0059] After multiple tests, the model classification accuracy can reach 100%.

[0060] Step 4: Use the best model to complete feature visualization.

[0061] Convolutional neural networks use convolutional layers and fully connected layers to extract and learn image features. In the convolutional layers of convolutional neural networks, each convolution kernel learns features with specific semantic information. The present invention uses this feature and uses Grad-CAM++ to obtain the importance of features related to the target category.

[0062] Grad-CAM++ is a method for explaining image classification models. It generates weights for feature map activations by calculating the sensitivity of the input image to the feature map. These weights are used to generate modified class activation images to highlight the areas that the model focuses on in image classification. It improves the activation images by introducing additional image-level gradient information and repeated forward and back propagation processes. Using Grad-CAM++, more reliable and accurate model explanations can be obtained, leading to a better understanding of the model's decision-making process. This method can also be applied to one-dimensional models and one-dimensional data.

[0063] After obtaining feature activation data through Grad-CAM++, the feature importance curve is drawn and intuitively visualized through heat map. There are two visualization methods: the individual visualization scheme randomly extracts eight samples from the data set and draws the feature importance curve respectively; the average visualization scheme calculates the feature importance of all samples in the data set, and then takes the average value of the data and feature importance in the sample for the entire data set to obtain the average sample, and finally draws the feature importance curve and heat map of the average sample.

[0064] Taking the negative ion data as an example, the effects of the individual visualization solutions proposed in step 4 are as follows: Figure 3 As shown, the average visualization scheme is as follows Figure 4 As shown in the figure. The horizontal axis represents the sequence number of the component; for the red curve, the vertical axis represents the normalized component content; for the blue curve, the vertical axis represents the feature importance, that is, the importance of the component corresponding to the sequence number in the model classification decision; the background image is a heat map of the feature importance curve, the darker the color, the higher the feature importance. It can be seen that the distribution of feature importance between different samples is stable and interpretable.

[0065] 3. Extract the top 10 metabolites.

[0066] The feature weights calculated by Grad-CAM++ were arranged from large to small, and the components with the highest diagnostic relevance were extracted. The corresponding markers were found according to the value of CompoundName (such as "0.70_1055.3143n"), and the top 10 markers extracted included. Through screening, the metabolites highly related to lymphoma were narrowed down from 17,182 to 10 (as shown in Tables 2 and 3 below).

[0067] (1) A combination of 10 serum metabolites: monoacylglycerol, phosphatidylserine, androgens and their derivatives, oxidized phosphoglycerates, steroid derivatives, steroids, glycerophospholipids, oxidized phosphoglycerate, medium-chain hydroxy acids and derivatives, and phosphatidylserine.

[0068] (2) A combination of 10 serum metabolites: phosphatidic acid, diglyceride, phosphatidic acid, decaglycerol, methionine, androgens, α-hydroxyprogesterone, testosterone acetate, dehydroepiandrosterone acetate, and methionine.

[0069] Table 2 Top ten metabolites detected by negative ion analysis in metabolomics .

[0070] Table 3 Top ten metabolites detected by positive ion detection in metabolomics .

[0071] 4. Use the XGBoost model to classify and verify the extracted markers.

[0072] 1) Negative ion data verification.

[0073] 120 negative ion data were used, including 58 negative samples (non-disease) and 62 positive samples (patients). Each data contains 10 features.

[0074] 80% of the data is randomly extracted as the training set, and the remaining 20% ​​is used as the test set. 20 tests are performed. The prediction accuracy is as follows: Figure 5 shown.

[0075] 2) Positive ion data verification.

[0076] 125 positive ion data are used, including 60 negative samples (non-disease) and 65 positive samples (patients). Each data contains 10 features.

[0077] 80% of the data was randomly extracted as the training set, and the remaining 20% ​​was used as the test set. Ten tests were performed, and the prediction accuracy was 100% (e.g. Figure 6 as shown).

[0078] In summary, the present invention provides a metabolomics data classification method, a model construction method, and a feature visualization method based on a one-dimensional neural network. After normalizing the one-dimensional metabolomics data, a cross-validation method is used to reduce the negative impact of a small data set size, determine the optimal hyperparameters and obtain a model, extract the features of the data through the optimal model, generate a curve graph using feature weights, and visualize the feature-based classification basis through a heat map, which is beneficial for doctors to further diagnose and analyze the data.

[0079] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An early auxiliary diagnostic marker for lymphoma, Features: The markers are the following (1) or / and (2): (1) a combination of 10 serum metabolites, monoacylglycerol, phosphatidylserine, androgens and their derivatives, oxidized phosphoglycerates, steroid derivatives, steroids, glycerophospholipids, oxidized phosphoglycerol phosphates, medium-chain hydroxy acids and derivatives, and phosphatidylserine; (2) A combination of 10 serum metabolites: phosphatidic acid, diglyceride, phosphatidic acid, decaglycerol, methionine, androgens, α-hydroxyprogesterone, testosterone acetate, dehydroepiandrosterone acetate, and methionine.

2. Use of the lymphoma early auxiliary diagnosis marker according to claim 1 in the preparation of lymphoma early auxiliary diagnosis products.

3. Use of a substance for detecting the lymphoma early auxiliary diagnosis marker according to claim 1 in the preparation of a lymphoma early auxiliary diagnosis product.

4. The use according to claim 3, Features: The substance is a substance used to detect the content of diagnostic markers in serum.

5. A kit for early auxiliary diagnosis of lymphoma, Features: The kit contains substances for detecting the early auxiliary diagnosis marker of lymphoma described in claim 1.

6. The early auxiliary diagnosis kit for lymphoma according to claim 5, Features: The kit also includes a carrier recording the model or a device loaded with the model.

7. A method for constructing an AI diagnostic model for lymphoma. Features: The following steps are involved: (1) Serum samples were measured by UPLC-Q / TOF MS to obtain serum metabolites; (2) The metabolites obtained in (1) were normalized and preprocessed to have zero mean and unit variance; (3) A simplified ResNet network is used for one-dimensional convolution calculation, and the loss function uses cross entropy loss; (4) Use K-Fold cross validation to obtain the best parameters and the best model for feature visualization; (5) After obtaining feature activation data using Grad-CAM++, plot feature importance curves and use heat maps; (6) Arrange the feature weights calculated by Grad-CAM++ from large to small according to the numerical value to obtain the marker described in claim 1.

8. An AI diagnostic model for lymphoma, It is characterized in that The lymphoma AI diagnostic model is constructed by the method for constructing the lymphoma AI diagnostic model as described in claim 7.

9. A computing device, It is characterized in that It comprises a memory and a processor, wherein the memory stores a disease screening model that can be loaded and executed by the processor to implement the method for constructing a lymphoma AI diagnostic model as described in claim 7.

10. A computer-readable storage medium, It is characterized in that The device stores a disease screening model that can be loaded and executed by a processor to implement the method for constructing a lymphoma AI diagnostic model as described in claim 7.

Images