Deep learning-based digital SERS kynurenine ultramicro detection method

Abstract

The invention discloses a digital SERS kynurenine ultramicro detection method based on deep learning. The method belongs to the technical field of biomedical detection and spectral analysis, and comprises the following steps: constructing an array type SERS substrate with uniformly distributed hot spots, and collecting multi-spatial-position SERS mapping spectrums of kynurenine samples with different concentrations; then performing unified cutting, abnormal value elimination, baseline correction, denoising and normalization processing on the spectral data, completing labeling by combining sample concentration, and constructing a standardized data set; and establishing a deep learning model fusing the improved ResNet network and a self-attention module, and jointly inputting the full-spectrum features and digital features obtained by feature peak dynamic threshold binaryzation to carry out concentration regression prediction. The method has high detection sensitivity, can realize quantitative detection of kynurenine as low as 10 <-11 > mol / L, has strong robustness to noise interference, and is suitable for detection and quantitative analysis of complex biological matrixes such as serum and urine.

Description

Technical Field

[0001] This invention relates to the field of biomedical detection and spectral analysis technology, specifically to a deep learning-based digital SERS method for ultra-micro detection of kynurenine. Background Technology

[0002] L-Kynurenine (L-Kyn) is an important intermediate in the tryptophan metabolic pathway, and its levels are closely related to immune activation, inflammatory responses, and various diseases such as Parkinson's disease, Alzheimer's disease, depression, and diabetes. The concentration of kynurenine in serum is usually extremely low. Traditional detection methods, such as high-performance liquid chromatography-mass spectrometry (HPLC-MS / MS) and enzyme-linked immunosorbent assay (ELISA), while possessing certain sensitivity, suffer from drawbacks such as expensive equipment, complex operation, long detection cycles, and poor reproducibility.

[0003] Surface-enhanced Raman scattering (SERS) technology has been increasingly applied to the detection of small molecule metabolites in recent years due to its high sensitivity and molecular fingerprint recognition capabilities. However, SERS detection still faces two major challenges in practical applications: first, the uneven distribution of hot spots on the substrate leads to large signal fluctuations at extremely low concentrations, limiting the stability of its quantitative analysis; second, traditional spectroscopic analysis methods rely on manual peak determination, which is difficult to handle complex background signals and low signal-to-noise ratio conditions.

[0004] Deep learning technology has demonstrated superior feature extraction and pattern recognition capabilities in spectral data modeling, enabling efficient identification and quantitative prediction of weak signals. Therefore, there is an urgent need for a novel detection method to achieve rapid and stable detection of kynurenine at extremely low concentration levels. Summary of the Invention

[0005] The present invention aims to solve the following technical problems:

[0006] (1) Improve signal distribution uniformity and robustness, and ensure signal repeatability and comparability at different locations;

[0007] (2) Utilize SERS mapping spectral data from multiple spatial locations instead of relying on a single spectrum to reduce random errors;

[0008] (3) Capture subtle spectral shape changes and characteristic peak intensity changes that are difficult to detect by traditional algorithms under extremely low concentration conditions;

[0009] (4) By learning the spatial correlation between spectra and making comprehensive use of the local and global features of mapping data, we can achieve higher sensitivity and stability in ultra-micro detection.

[0010] To achieve the above objectives, this invention proposes a deep learning-based digital SERS method for ultra-micro detection of kynurenine, comprising the following steps:

[0011] Step 1: Prepare a highly sensitive, uniform array-type SERS substrate.

[0012] A monolayer polystyrene (PS) microsphere array was fabricated on a silicon wafer surface using a self-assembly technique. Subsequently, reactive ion etching (RIE) was employed to form a conical silicon nanostructure at the PS sphere-silicon interface. Further, Au / Ag noble metal thin films were deposited via magnetron sputtering to form a uniformly distributed array of nano-hot spots. This substrate ensures the spatial uniformity of the electromagnetic field-enhanced hot spots, reducing signal differences at different locations and significantly improving the signal-to-noise ratio (SNR) and detection sensitivity.

[0013] Step 2: Sample spectral acquisition. For concentrations in C ∈ [10...],... -11 10 -3 Different concentrations of kynurenine aqueous solution in the mol / L range were dropped onto the surface of a SERS substrate. Three parallel samples were prepared for each concentration, using 2 µL of solution each time. After drying, Raman signals were acquired. Raman spectrometers were used to collect n rows of data for each concentration sample. n columns = m SERS mapping signals from different spatial locations, resulting in a dataset. ,in The Raman shift is used; preferably, the detection conditions are: laser wavelength 785nm, power 20mW, integration time 500ms.

[0014] Step 3: Batch preprocessing of spectral data. Perform the following processing on each spectrum in the dataset:

[0015] 3.1 Spectral cropping: A uniform spectral cutoff of 500-1800 cm⁻¹ should be maintained. -1 Raman signals within the range:

[0016] = v∈[500,1800],

[0017] 3.2 Outlier Removal: The Whitaker-Hayes algorithm is used to remove outlier peaks.

[0018] =RemoveOutlier( ),

[0019] 3.3 Baseline Correction: Baseline correction is performed using the Asymmetric Penalized Least Squares (ASPLS) algorithm.

[0020] = -Baseline ASPLS ,

[0021] 3.4 Noise Reduction: Smoothing is achieved by applying a Savitzky-Golay filter.

[0022] =SavGol( ,w,p),

[0023] 3.5 Normalization: Vector normalization is used to make different spectral intensities comparable.

[0024] = ,

[0025] Step 4: Digitize the spectral signal. For each preprocessed spectrum, the characteristic peak position of kynurenine is v0 = 560 ± 1 cm. -1 Signal strength detected at the location:

[0026] B i = ,

[0027] Where T is the dynamic threshold, which can be defined as:

[0028] ,

[0029] μ bg σ is the average intensity of the background region. bg denoted as the standard deviation of the background area intensity, and k is an empirical coefficient (usually k=5).

[0030] The digital positive rate (RPV) is obtained by analyzing m spectra collected for each concentration sample.

[0031] ,

[0032] RPV, as a concentration-related statistical indicator, is input into the deep learning model as an auxiliary feature.

[0033] Step 5: Deep Learning Modeling. Construct a deep learning model based on the fusion of convolutional neural networks and attention mechanisms:

[0034] 5.1 The backbone network adopts an improved ResNet:

[0035] F=ResNet( ),

[0036] 5.2. Introducing multi-layer convolution and self-attention modules (MHSA) to capture global semantics across the spectrum:

[0037] F att =MHSA(F),

[0038] 5.3. Feature fusion is achieved through a fully connected layer, integrating digital statistical features:

[0039] F fusion =Concat(F att (RPV)

[0040] 5.4 Output Concentration Prediction Value:

[0041] =MLP(F fusion ),

[0042] The loss function adopts the fusion loss form:

[0043] ,

[0044] Where λ1 and λ2 are weighting coefficients, MSE is the mean squared error, and BCE is the binary cross-entropy. V represents the predicted digital statistical feature, and RPV represents the actual digital statistical feature, ensuring consistency between the predicted concentration and the digital feature statistics.

[0045] Step 6: Model Training and Validation. Data is divided into three sets: 80% training set, 10% validation set, and 10% test set. Preferably, the Adam optimizer is used with a learning rate of 1×10⁻⁶. -3 Weight decay 1×10 -4 Batch size is 32, training duration is 200 epochs; early stopping strategy is used to avoid overfitting, and the coefficient of determination R0 is [value missing]. 2 Mean squared error (MSE) and mean absolute error (MAE) are used as performance evaluation indicators.

[0046] Step 7: Quantitative Regression and Result Output. Input the spectral set to be measured {Si}, and the model outputs the predicted concentration. Simultaneously, output R0. 2 The confidence interval is predicted, ultimately achieving highly sensitive, repeatable, and automated quantitative detection of ultra-low concentrations of kynurenine.

[0047] Beneficial effects:

[0048] Compared with existing detection methods, the present invention has the following significant advantages and technical effects:

[0049] I. Significantly Improved Sensitivity: This invention employs multi-point mapping acquisition and statistical modeling, avoiding random errors caused by relying on a single spectrum, and can detect values ​​as low as 10. -11 The concentration of kynurenine in mol / L is more accurate than that of traditional methods.

[0050] II. Strong ability to capture weak signals: This invention introduces multi-layer convolution and self-attention mechanism, which can extract subtle differences in spectral shape and global correlation across the spectrum, improve the ability to identify weak feature peaks, and still output prediction results stably even when the signal-to-noise ratio is low.

[0051] 3. Digital Feature-Assisted Regression: SERS spectra are mapped into binary digital features, which are then combined with deep learning regression models to improve the physical consistency of prediction results and model robustness, effectively reducing overfitting.

[0052] IV. Joint Loss Optimization: By jointly optimizing regression loss and statistical feature constraints, the predicted concentration not only approximates the true value, but also maintains consistency with the statistical characteristics of RPV, thereby improving the model's generalization ability and interpretability.

[0053] V. Highly automated and scalable: The entire process of preprocessing, feature extraction, and model prediction is automated, and it can be extended to the detection of ultra-micro samples of other similar samples, enabling simultaneous quantitative analysis of multiple indicators.

[0054] VI. High clinical application potential: This method has low sensitivity to environmental interference and batch-to-batch differences, and is suitable for highly sensitive detection of kynurenine in complex media such as serum and urine. Attached Figure Description

[0055] Figure 1 This is a flowchart of the detection method described in this invention; Figure 2 The image shows the SERS spectrum of kynurenine collected in Specific Example 1; Figure 3 This is a linear relationship graph between the predicted concentration and the actual concentration in Specific Example 1; Figure 4 The image shows the SERS spectrum of kynurenine collected in Specific Example 2; Figure 5 This is a graph showing the linear relationship between the predicted concentration and the actual concentration in Specific Example 2. Detailed Implementation

[0056] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. However, the following embodiments are only for explaining the present invention, and the scope of protection of the present invention should include all the contents of the claims. Moreover, through the description of the following embodiments, those skilled in the art can fully implement all the contents of the claims of the present invention.

[0057] Example 1:

[0058] This embodiment provides a deep learning-based digital SERS method for ultra-micro detection of kynurenine, the process of which is as follows: Figure 1 As shown, it includes the following steps:

[0059] Step 1: Prepare a highly sensitive array-type SERS substrate.

[0060] A monocrystalline silicon wafer was selected as the substrate, and a monolayer polystyrene (PS) microsphere array was formed using interface self-assembly technology. Reactive ion etching (RIE) was then used to form periodic tapered silicon nanostructures at the interface between the PS spheres and the silicon wafer. After etching, Au / Ag noble metals were deposited using physical sputtering technology to form a uniformly distributed metal nanoarray.

[0061] Step 2: Sample preparation and spectral acquisition.

[0062] Preparation of C={10 -11 10 -10 10 -9 10 -8 10 -7 10 -6 10 -5 10 -4 A 1 mol / L kynurenine aqueous solution was prepared, with three parallel samples for each concentration, and each drop volume was 2 μL. After drying, the sample was placed under a Raman spectrometer to collect SERS spectra. The acquisition conditions were: laser wavelength 785 nm, laser power 20 mW, integration time 500 ms. 400 mapping points were collected for each sample, resulting in a spectral set of 9600 spectra.

[0063] ,

[0064] Where, v∈[0,2000], C={10 -11 10 -10 10 -9 10 -8 10 -7 10 -6 10 -5 10 -4}, and perform concentration labeling on the dataset.

[0065] Step 3: Batch preprocessing of spectral data. Perform the following processing on each spectrum in the dataset:

[0066] 3.1 Spectral cropping: A uniform spectral cutoff of 500-1800 cm⁻¹ should be maintained. -1 Interval signal.

[0067] = v∈[500,1800],

[0068] 3.2 Outlier Removal: The Whitaker-Hayes algorithm is used to remove outliers.

[0069] =RemoveOutlier( ),

[0070] 3.3 Baseline Correction: Baseline correction is performed using the Asymmetric Penalized Least Squares (ASPLS) algorithm.

[0071] = -BaselineASPLS ,

[0072] 3.4 Noise Reduction and Smoothing: The Savitzky-Golay filter is used for smoothing. The window width w and the polynomial order p are determined according to the noise level.

[0073] =SavGol( ,w,p),

[0074] 3.5 Normalization: Perform vector normalization to obtain the normalized spectrum.

[0075] = ,

[0076] Step 4: Digitize the spectral signal. For each preprocessed spectrum, the characteristic peak position of kynurenine is ν0 = 560 ± 1 cm. -1 Signal strength detected at the location:

[0077] Bi= ,

[0078] Where T is the dynamic threshold, which can be defined as:

[0079] ,

[0080] μ bg σ bg , respectively, represent the mean and standard deviation of the background region intensity, and k is an empirical coefficient (k=5).

[0081] The digital positive rate (RPV) was obtained by analyzing all 400 spectra:

[0082] ,

[0083] RPV, as a concentration-related statistical indicator, is input into the deep learning model as an auxiliary feature.

[0084] Step 5: Deep Learning Modeling. Construct a deep learning model based on the fusion of convolutional neural networks and attention mechanisms:

[0085] 5.1 The backbone network adopts an improved ResNet:

[0086] F=ResNet( ),

[0087] 5.2. Introducing multi-layer convolution and self-attention modules (MHSA) to capture global semantics across the spectrum:

[0088] F att =MHSA(F),

[0089] 5.3. Feature fusion is achieved through a fully connected layer, integrating digital statistical features:

[0090] F fusion =Concat(F att (RPV)

[0091] 5.4 Output Concentration Prediction Value :

[0092] =MLP(F fusion ),

[0093] The loss function adopts the fusion loss form:

[0094] ,

[0095] Where λ1 and λ2 are weighting coefficients, MSE is mean squared error, and BCE is binary cross-entropy, ensuring consistency between predicted concentration and digitized feature statistics.

[0096] Step 6: Model Training and Validation. Data is divided into three sets: 80% training, 10% validation, and 10% test. The Adam optimizer is used with a learning rate of 1×10⁻⁶. -3 Weight decay 1×10 -4 Batch size is 32, training duration is 200 epochs; early stopping strategy is used to avoid overfitting, and the coefficient of determination R0 is [value missing]. 2 Mean squared error (MSE) and mean absolute error (MAE) are used as performance evaluation indicators.

[0097] Step 7: Quantitative Regression and Result Output. Input the spectral set to be measured {Si}, and the model outputs the predicted concentration. Simultaneously output R 2 The confidence interval is predicted, ultimately achieving highly sensitive, repeatable, and automated quantitative detection of ultra-low concentrations of kynurenine.

[0098] The kynurenine spectral data collected in step 2 of this embodiment are as follows: Figure 2 As shown.

[0099] This implementation method is in C={10 -11 10 -10 10 -9 10 -8 10 -7 10 -6 10 -5 10 -4 At eight concentration gradients of 1 mol / L, 400 spectra were collected for each concentration sample, and the experiment was repeated three times, resulting in a total of 9600 spectra. 7680 of these were used as the training set, and 960 as the validation set. The deep learning model was then trained and optimized following the steps outlined above.

[0100] The trained model was then used to evaluate its performance on the remaining 960 test set spectra, yielding the following concentration prediction results: Figure 3 As shown, the coefficient of determination R 2 =0.993, indicating that the model has high prediction accuracy.

[0101] Example 2 The same experimental testing was performed on patient serum samples using the same method as in Example 1. The difference from Example 1 is that step 2 was replaced; the remaining steps are the same as in Example 1. Step 2 of this example is as follows: Step 2: Patient serum sample pretreatment and spectral acquisition 2.1. Serum pretreatment To address the complex background interference (such as proteins, lipids, and carbohydrates) in patient serum samples, the following pretreatment method was adopted: 100 µL of patient serum sample was taken, 300 µL of acetonitrile was added, and the mixture was thoroughly mixed on a vortex mixer and centrifuged at 13000 rpm for 10 minutes. The supernatant was collected for later use. The supernatant was then passed through a pretreated C18 extraction column (activated with 1 mL of methanol and then equilibrated with 1 mL of ultrapure water). Methanol and water were used as the elution buffer at a ratio of 5:95 (containing 0.1% formic acid), and the target component solution was collected.

[0102] 2.2. SERS Spectral Acquisition Serum samples were collected from 10 patients with radiation enteritis. The sample solutions, processed as described above, were dropped onto the surface of a SERS substrate. Three parallel samples were prepared for each concentration, using 2 µL of solution each time. After drying, Raman signals were acquired. The acquisition conditions were: laser wavelength 785 nm, power 20 mW, integration time 500 ms. 400 mapping points were acquired for each sample, resulting in a spectral set of 12,000 lines. , And the dataset is labeled with concentration. In step 2 of this implementation method, serum sample spectral data from 10 patients are collected as follows: Figure 4 As shown.

[0103] In this implementation method, 400 spectra were collected from each patient's serum sample, and the experiment was repeated three times, resulting in a total of 12,000 spectra. 9,600 of these were used as the training set, and 1,200 as the validation set. The deep learning model was then trained and optimized according to the aforementioned steps. The trained model was then used to evaluate its performance on the remaining 1200 test set spectra, yielding the concentration prediction results as follows: Figure 5 As shown, the coefficient of determination R² = 0.981, indicating that the model has high predictive accuracy.

[0104] Through the above embodiments, the present invention achieves the following technical effects:

[0105] 1. The fabricated array-type SERS substrate exhibits highly uniform hotspot distribution, significantly improving signal repeatability and spatial consistency;

[0106] 2. By utilizing the statistical characteristics of multi-position mapping spectra, the signal reliability at low concentrations is effectively improved, achieving 10 -11 Detection limits at the mol / L level;

[0107] 3. By combining an improved ResNet with a self-attention mechanism, the ability to capture weak feature peaks is enhanced, and the robustness of the model to noise is improved.

[0108] 4. The method automates the entire process from spectral acquisition and preprocessing to concentration prediction. It is suitable for early clinical detection and quantitative analysis of kynurenine in complex biological matrices such as serum and urine, and can be extended to trace detection of other target molecules.

[0109] The above description is merely a specific embodiment of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A deep learning-based digital SERS method for ultra-micro detection of kynurenine, characterized in that, Includes the following steps: Step 1: Prepare a highly sensitive, uniform array-type SERS substrate; Step 2: Raman spectra of kynurenine samples of different concentrations were acquired using a SERS array substrate to obtain mapping spectral data containing multiple spatial locations for each concentration sample; Step 3: Perform batch preprocessing on the spectra: spectral cropping, outlier removal, baseline correction, noise reduction, normalization, and concentration labeling; Step 4: Binarize each preprocessed spectrum and statistically analyze all concentration samples to obtain digital features; Step 5: Construct a deep learning model based on the fusion of convolutional neural network and attention mechanism to extract spectral features, stitch together the digital features obtained in step 4, and achieve feature fusion through a fully connected layer; Step 6: Use the collected data to train and validate the deep learning model, and optimize it using a fusion loss function; Step 7: Input the spectral set to be measured, the model outputs the predicted concentration, and also outputs the coefficient of determination R. 2 And the predicted confidence interval; This method is suitable for the ultra-micro detection of kynurenine in serum or urine matrices.

2. The method according to claim 1, characterized in that, The preparation of the SERS substrate in step 1 specifically includes: preparing a monolayer polystyrene microsphere array on the surface of a silicon wafer using self-assembly technology; forming a conical silicon nanostructure at the interface between the PS spheres and silicon using reactive ion etching; and depositing an Au / Ag noble metal thin film by magnetron sputtering to form a uniformly distributed array substrate.

3. The method according to claim 1, characterized in that, The detection limit for the concentration acquired by Raman spectroscopy in step 2 is as low as 10. -11 The Raman spectroscopy acquisition conditions were as follows: laser wavelength 785 nm, power 15-25 mW, integration time 500 ms, and spectral range 0-2000 cm⁻¹. -1 .

4. The method according to claim 1, characterized in that, Step 3, the batch spectral preprocessing steps, specifically include: Spectral uniform clipping retains 500-1800 cm⁻¹ -1 Raman signals within the range were analyzed; outliers were removed using the Whitaker-Hayes algorithm; baseline removal was performed using an asymmetric penalized least squares algorithm; smoothing was performed using a Savitzky-Golay filter with a window size of 5-21 and a polynomial order of 2-4; and vector normalization was used to make the peak characteristics of different spectra directly comparable.

5. The method according to claim 1, characterized in that, The dynamic threshold T for binarization in step 4 is defined as follows: ， Where, μ bg σ is the mean intensity of the background region. bg denoted as the standard deviation of the background area intensity, and k is an empirical coefficient with a value of 4-6.

6. The method according to claim 1, characterized in that, The deep learning model mentioned in step 5 includes: The backbone network adopts an improved ResNet architecture; multi-layer convolution and self-attention modules are introduced to capture global semantics across the spectrum; digital statistical features (RPV) are fused with deep learning features; and concentration prediction values ​​are output through a multi-layer perceptron.

7. The method according to claim 6, characterized in that, The fusion loss function is: ， Where λ1 and λ2 are weighting coefficients, MSE is the mean squared error loss, and BCE is the binary cross-entropy loss. To predict the concentration, C represents the actual concentration. V represents the predicted digital statistical feature, and RPV represents the actual digital statistical feature.

8. The method according to claim 1, characterized in that, The deep learning model described in step 6 is trained using the Adam optimizer with a learning rate of 1×10⁻⁶. -3 Up to 1×10 -4 The weight decays to 1×10 -4 Up to 1×10 -5 The batch size is 16-64, and the number of training rounds is 150-250.

9. The method according to claim 1, characterized in that, The training of the deep learning model described in step 6 employs an early stopping strategy to avoid overfitting, with a determination coefficient R0. 2 Mean square error and mean absolute error are used as performance evaluation indicators.

Images