SERS-based chiral target detection method, system and application thereof

Abstract

The application discloses a kind of chiral target detection method, system and application based on SERS.Therein, the method includes: chiral gold nanoparticles are mixed with the sample to be measured containing chiral target substance in liquid phase system and are incubated, and mixture is obtained;The mixture is applied to the surface of enhanced raman spectrum detection substrate, and SERS detection is carried out, and the detection of chiral target substance is realized;Wherein chiral gold nanoparticles and chiral target substance have the same chiral configuration;Enhanced raman spectrum detection substrate includes glass fiber membrane and three-dimensional porous network structure combined on glass fiber membrane, and the three-dimensional porous network structure is formed by silver nanoparticle accumulation.The application shows good recognition ability to a variety of chiral amino acids, especially can realize high-precision quantitative detection of D-serine, can be widely applied to rapid, non-invasive auxiliary screening of depression and Alzheimer's disease related biomarkers, and provides a new technical path for real-time detection of chiral molecules.

Description

Technical Field

[0001] This invention relates to the field of chemical analysis technology, and particularly to a method, system, and application for detecting chiral target substances based on a chiral gold nanoparticle-silver modified glass fiber surface-enhanced Raman substrate. Background Technology

[0002] Amino acids are essential building blocks of living systems and key molecules in proteins, enzymes, neurotransmitters, and various metabolic pathways. Except for glycine, most amino acids possess a chiral center and exist as L- and D-enantiomers. Although the two enantiomers are highly similar in molecular composition and most conventional physicochemical parameters, their different spatial configurations often lead to significant differences in their recognition processes, receptor binding behavior, metabolic pathways, and physiological functions within organisms. Therefore, the recognition and quantitative analysis of chiral molecules has always been an important research direction in analytical chemistry, biomedicine, and molecular diagnostics. D-serine has attracted widespread attention due to its close relationship with nervous system function. Existing research shows that D-serine is closely related to neuropsychiatric diseases such as depression and Alzheimer's disease; its level in cerebrospinal fluid is negatively correlated with the degree of depression, and changes in serum levels also have potential biomarker value. Therefore, establishing a rapid, sensitive, and accurate method for detecting D-serine in complex sample environments is not only of significant basic research importance, but also has the potential for application in assisting screening and immediate detection of related biomarkers such as depression and Alzheimer's disease. Currently, analytical methods for chiral molecules mainly include nuclear magnetic resonance (NMR), circular dichroism spectroscopy (CD), circularly polarized luminescence (EPL), Raman spectroscopy, mass spectrometry (MS), colorimetry, and enzymatic methods. While each method has its advantages in experimental research, they generally suffer from drawbacks such as expensive equipment, complex operation, long analysis cycles, demanding pretreatment requirements, insufficient adaptability to complex biological matrices, and limited field applications.

[0003] In recent years, surface-enhanced Raman scattering (SERS) technology has attracted widespread attention in the detection of trace biomolecules due to its advantages such as high sensitivity, fast response, and distinct fingerprint spectral characteristics. Meanwhile, the local electromagnetic field enhancement effect based on chiral plasmon nanostructures provides a new technical path for the optical recognition of chiral molecules. However, existing chiral SERS detection systems still have certain limitations: firstly, most traditional substrates are predominantly two-dimensional structures with limited and insufficiently uniform hotspot distribution, leading to unstable signal enhancement and poor repeatability; secondly, in complex systems, Raman spectra are easily affected by baseline drift, background scattering, overlapping characteristic peaks, and interference from outliers, resulting in decreased accuracy in quantitative analysis.

[0004] In contrast, three-dimensional porous substrates possess higher specific surface area and more abundant electromagnetic coupling sites, which is beneficial for constructing high-density, stable "hot spot" regions, thereby enhancing Raman signal enhancement. However, existing research still lacks a systematic explanation of the molecular-nano interface interaction mechanism during chiral recognition, making it difficult to explain the differences in adsorption behavior of different amino acid enantiomers on nanosurfaces. Furthermore, for the quantitative analysis of complex spectral data, methods relying solely on manual peak selection or single characteristic ratios are insufficient to effectively address issues of multi-peak overlap and noise interference.

[0005] Therefore, high-precision quantitative analysis of complex SERS spectra typically requires a combination of data-driven methods, including standardization, feature reduction, and information filtering of the raw spectra, and improvement of prediction stability through multi-model fusion or machine learning methods. However, how to effectively integrate chiral plasmon nanorecognition mechanisms, highly enhanced SERS substrates, and intelligent spectral analysis methods to construct a detection system with both high selectivity and robust quantitative analysis capabilities remains a key technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] The main objective of this invention is to provide a chiral target detection method, system, and application based on SERS, so as to overcome the technical problems existing in the prior art.

[0007] To achieve the above-mentioned objectives, the present invention provides the following technical solution.

[0008] According to a first aspect of the present invention, a method for preparing chiral gold nanoparticles is provided, comprising: A gold octahedral seed crystal is provided, wherein the surface of the gold octahedral seed crystal is entirely composed of (111) crystal planes; Using a chiral dipeptide as a chiral template, gold atoms are induced to be deposited in a directional manner on the surface of the gold octahedral seed crystal, thereby forming the chiral gold nanoparticles.

[0009] In one embodiment, the preparation method may specifically include: adding an alkali to an aqueous solution containing 0.1-0.5 mM chloroauric acid, 10-100 mM ascorbic acid and 1-10 mM CTAB at a temperature of 25-60°C to react and obtain the gold octahedral seed crystal, wherein the molar ratio of chloroauric acid to alkali is 0.5-2.5.

[0010] In one embodiment, the preparation method may specifically include: adding ascorbic acid to an aqueous solution containing 0.1-0.5 mM chloroauric acid and 1-10 mM CTAB at room temperature, and maintaining a molar ratio of chloroauric acid to ascorbic acid of 0.5-2.5, thereby obtaining a product containing Au. +The growth solution is prepared by adding a chiral dipeptide and the gold octahedral seed crystals to the growth solution and mixing them evenly to make the concentration of the chiral dipeptide 0.5~4 mM and the concentration of the gold octahedral seed crystals 0.2~1 D. The mixture is then reacted for more than 1 hour to obtain the chiral gold nanoparticles.

[0011] In one embodiment, the chiral dipeptide comprises a cysteine-leucine dipeptide. When L-type and D-type cysteine-leucine dipeptides are used as templates for the synthesis of chiral gold nanoparticles, respectively, the strong interaction between the cysteine ​​thiol groups and the surface of the gold octahedral seed crystals forms Au-S bonds, inducing the directional deposition of gold atoms along specific crystal planes, thereby obtaining L-type and D-type gold nanoparticles with mirror-symmetric chirality. These chiral gold nanoparticles exhibit a mirror-symmetric response on circular dichroism spectroscopy, while conventional UV-Vis absorption cannot directly distinguish between the two.

[0012] According to a second aspect of the present invention, a chiral gold nanoparticle is provided, which is prepared by any of the chiral gold nanoparticle preparation methods described in the present invention.

[0013] Furthermore, by utilizing the electromagnetic adsorption between the chiral gold nanoparticles and the chiral target material with the same chiral configuration, the surface-enhanced Raman spectral signal of the chiral target material can be effectively enhanced. Consequently, the concentration of the chiral target material can be quantitatively detected based on changes in the surface-enhanced Raman spectral signal.

[0014] Taking chiral amino acid molecules as the chiral target material as an example, the recognition between the target chiral amino acid molecules and chiral gold nanoparticles is mainly based on three interaction types: first, sulfur-containing amino acids achieve strong adsorption through Au-S bonds; second, amino acids with aromatic ring conjugated structures achieve enrichment through Au-π interactions; and third, some amino acids without thiol groups and large π conjugated structures achieve chiral selective adsorption through chiral stereomatching and electrostatic interactions. Among these, amino acids that can be effectively distinguished and have enhanced detection include cysteine, histidine, methionine, serine, threonine, tryptophan, and tyrosine.

[0015] For example, if the chiral target substance is chiral serine, the Raman detection light source wavelength can be set to 785 nm, the power to 100 mW, and the integration time to 5 s, and surface-enhanced Raman spectroscopy detection can be performed using corresponding chiral gold nanoparticles. Of course, if the chiral target substance is other amino acids or other compounds, the corresponding Raman detection light source wavelength, power, and integration time should be adjusted appropriately, and are not limited to the scope exemplified in this invention.

[0016] According to a third aspect of the present invention, a method for preparing an enhanced Raman spectroscopy detection substrate is provided, comprising: depositing silver nanoparticles on a glass fiber membrane by a silver mirror reaction, thereby forming the enhanced Raman spectroscopy detection substrate.

[0017] In one embodiment, the preparation method specifically includes: fully immersing a glass fiber membrane in a silver ammonia solution, removing it, rinsing it with deionized water, then immersing it in a reducing solution and incubating it at 10-60°C for 25-60 min to obtain the enhanced Raman spectroscopy detection substrate; wherein, the concentration of the silver ammonia solution, calculated as AgNO3, is 0.5-1%; and the reducing solution is an aqueous solution containing 50-200 mM glucose. The size of the silver nanoparticles can be adjusted by changing the concentrations of silver ions and glucose.

[0018] According to a fourth aspect of the present invention, an enhanced Raman spectroscopy detection substrate is provided, comprising a glass fiber membrane and a three-dimensional porous network structure bonded to the glass fiber membrane, the three-dimensional porous network structure being formed by the stacking of silver nanoparticles.

[0019] Furthermore, the enhanced Raman spectroscopy detection substrate is prepared by any of the enhanced Raman spectroscopy detection substrate preparation methods described in this invention.

[0020] Furthermore, the enhanced Raman spectroscopy detection substrate has a three-dimensional porous interconnected network, which can significantly increase the specific surface area and form a high-density SERS hot spot. When chiral gold nanoparticles adsorbed with the target chiral material are deposited on the surface of the substrate, bimetallic plasmon coupling can further occur, thereby significantly enhancing the local electromagnetic field and improving the SERS detection sensitivity.

[0021] According to a fifth aspect of the present invention, a spectral data intelligent analysis module is provided for preprocessing, feature screening and concentration prediction of surface-enhanced Raman spectroscopy detection data of a test sample, thereby realizing quantitative analysis of chiral target substances in the test sample.

[0022] The intelligent spectral data analysis module includes a trained quantitative prediction model for chiral target substances. This model is used to simultaneously predict the concentrations of D-type and L-type chiral target substances based on surface-enhanced Raman spectroscopy data.

[0023] The quantitative prediction model was trained using surface-enhanced Raman spectroscopy sample data with known concentrations of D-type and L-type chiral target substances. During training, partial least squares was first used to extract latent variables from the high-dimensional Raman spectral data under supervision, and cross-validation was used to determine the optimal number of latent variables to obtain low-dimensional latent variable features related to changes in the concentrations of D-type and L-type chiral target substances.

[0024] The low-dimensional latent variable features are input into a one-dimensional convolutional-bidirectional long short-term memory feature reconstruction unit. This unit comprises a one-dimensional convolutional layer, a batch normalization layer, a Gaussian error linear unit activation layer, and a bidirectional long short-term memory neural network layer. The one-dimensional convolutional layer extracts local combined features between adjacent latent variables, while the bidirectional long short-term memory neural network layer extracts forward and backward dependencies in the latent variable sequence, thereby obtaining a deep spectral characterization.

[0025] The deep spectral characterization is further input into a Bayesian approximate neural network prediction unit. This prediction unit comprises, in sequence, a first fully connected layer, a batch normalization layer, a Gaussian error linear unit activation layer, a first random deactivation layer, a second fully connected layer, another Gaussian error linear unit activation layer, a second random deactivation layer, and an output layer. The output layer includes two output nodes, corresponding to the concentrations of D-type chiral target substances and L-type chiral target substances, respectively.

[0026] During model training, an improved dual-output weighted robust constraint loss function is used to optimize the joint quantitative prediction model. Compared to the traditional mean squared error loss function, which only penalizes the squared error between the predicted and true values, the improved dual-output weighted robust constraint loss function has the following improvements: First, it employs D-type and L-type dual-output error terms, enabling the model to simultaneously learn the concentration response relationship of two enantiomers; second, it replaces the simple squared error term with a Huber robust error term, maintaining high sensitivity within a small error range and reducing the interference of anomalous spectra on parameter updates within a large error range; third, it introduces a model parameter regularization term to suppress overfitting and improve the model's generalization ability to unknown samples. During training, the parameters of the quantitative prediction model are updated through backpropagation, and the final model parameters are determined using a weight decay optimization strategy, a learning rate scheduling strategy, and an early stopping strategy. Therefore, the quantitative prediction model can extract effective features related to the concentration of chiral enantiomers from complex surface-enhanced Raman spectroscopy signals and achieve quantitative analysis of D-type and L-type chiral target substances. For quantitative analysis in complex systems, Raman spectroscopy data is easily affected by baseline drift, scattering effects, characteristic peak overlap, matrix differences, and anomalous spectral points. Therefore, relying solely on manual peak selection, single peak intensity ratios, or empirical interpretation is insufficient to meet the high-precision quantitative detection requirements of chiral target substances in complex samples. The intelligent spectral data analysis module in this embodiment includes a spectral preprocessing unit, a partial least squares (PLS) latent variable extraction unit, a one-dimensional convolutional neural network-bidirectional long short-term memory (LSTM) feature reconstruction unit, and a Bayesian approximate neural network prediction unit based on Monte Carlo random inactivation. This module first performs standard normal variable (SNV) transformation and variable standardization on the Raman spectrum to reduce the impact of overall spectral intensity fluctuations and differences in sample testing conditions. Then, it uses the PLS algorithm to extract low-dimensional latent variable features related to the concentrations of D-serine and L-serine. Next, it uses a one-dimensional convolutional neural network-bidirectional LSTM to nonlinearly reconstruct the latent variable features. Finally, it outputs the predicted concentrations of D-serine and L-serine through a Bayesian approximation neural network and obtains the prediction uncertainty based on multiple Monte Carlo forward propagation, thereby improving the reliability and accuracy of quantitative analysis of chiral target substances under complex matrix conditions.

[0027] The input to the intelligent spectral data analysis module is the Raman spectral vector of the sample to be tested: in, Indicates the first The sample in the first Spectral intensity at each Raman shift This represents the number of spectral sampling points. The model output is: in, This indicates the output prediction result. This indicates the predicted concentration of D-serine. Indicates the predicted concentration of L-serine The preprocessing stage first performs SNV on the raw Raman spectrum to reduce the impact of overall spectral intensity fluctuations, substrate differences, and variations in sample testing conditions. For the first... The spectrum in the first The intensity at each Raman shift is processed using the SNV method as follows: in, For the first The average intensity of the spectrum, For the first Standard deviation of the spectrum To prevent extremely small constants from being divided by zero, in one specific embodiment, Take 1 10 -8 .

[0028] Furthermore, to ensure that the spectral variables at different Raman shifts are on the same scale, after performing the SNV transformation, the mean and standard deviation of each Raman shift variable in the training set are calculated, and standardized as follows: in, Indicates the first The spectrum in the first The spectral intensity at each Raman shift after SNV processing Indicates the first training set The average value of each Raman displacement variable, Indicates the first training set The standard deviations of the Raman shift variables. These standardized parameters are calculated only from the training set and used for a unified transformation of the training set, validation set, test set, and the samples to be predicted. Through this process, each Raman shift variable is transformed into a dimensionless variable centered on the mean of the training set and scaled by the standard deviation of the training set, thus placing different variables on the same numerical scale and avoiding excessive weighting of high-intensity variables in model training.

[0029] The PLS feature extraction unit is implemented using an existing PLS algorithm, preferably the NIPALS or SIMPLS algorithm, but other equivalent PLS algorithms can also be used. This unit is not a newly constructed neural network layer, but rather an existing multivariate statistical regression feature extraction method. Its input is the preprocessed spectral matrix. and the corresponding concentration label matrix ,in , The number of training samples, This represents the number of spectral sampling points. The two columns represent the actual concentrations of D-serine and L-serine, respectively.

[0030] The PLS feature extraction unit searches for features that can simultaneously interpret the spectral matrix. Change and concentration label matrix The changing direction of the latent variables projects the high-dimensional spectral variables into low-dimensional latent variable scores. Specifically, in the first... During the extraction of latent variables, the PLS algorithm determines the weight vector. The score vector obtained from the spectral matrix is: With concentration label matrix There is a large covariance between them. Then, based on the score vector... Calculate the corresponding loading vectors and update the residuals of the spectral matrix and concentration label matrix to extract the next latent variable. Repeat the above process to obtain... Scores of each latent variable: in, This serves as the input to the subsequent one-dimensional convolutional neural network - LSTM feature reconstruction module. Through this step, the original high-dimensional Raman spectrum is converted into low-dimensional features related to the concentrations of D-serine and L-serine, thereby reducing the impact of spectral noise, variable collinearity, and feature peak overlap on subsequent neural network training.

[0031] The number of latent variables in PLS The number of candidate latent variables was determined using five-fold cross-validation. Specifically, the training samples with true concentration labels were randomly divided into five subsets. Each time, four subsets were used as the modeling set, and the remaining subset as the validation set; this process was repeated, ensuring that each subset served as the validation set at least once. PLS models were established separately, and the predicted concentrations of all training samples during cross-validation were obtained. Then, the root mean square error of cross-validation for D-serine and L-serine was calculated separately, and the optimal number of latent variables was determined based on minimizing the sum of the two. The root mean square error of the cross-validation is calculated as follows: in, To be the optimal number of latent variables, Indicates the first The true concentration of each sample Indicates the first The predicted concentration obtained when each sample is used as a validation sample in cross-validation. Number of PLS ​​latent variables. These are model hyperparameters, determined through cross-validation before training the neural network weights, and are not part of the neural network weight parameters updated through backpropagation. Subsequently, the PLS model is rebuilt using all training samples, and both the training samples and the samples to be predicted are converted into PLS latent variable features of the same dimension. In one specific embodiment, the labeled spectral samples are stratified and randomly divided according to concentration combinations, with 80% used as the training set and 20% as the test set; the training set is further divided into a validation set, or five-fold cross-validation is used to determine the number of PLS ​​latent variables and some hyperparameters. The test set is not involved in the determination of SNV normalization parameters, PLS model, neural network parameters, and hyperparameters.

[0032] The one-dimensional convolutional neural network-LSTM feature reconstruction module is used to receive the PLS latent variable features. In this embodiment, the input to the one-dimensional convolutional layer is the spectral feature vector after preprocessing and PLS dimensionality reduction. As an alternative implementation, without setting a PLS feature extraction unit, the preprocessed complete Raman spectral vector can also be directly input into the one-dimensional convolutional layer.

[0033] The feature reconstruction module comprises a one-dimensional convolutional layer, a batch normalization layer, a Gaussian error linear unit activation function layer, and a bidirectional long short-term memory network layer connected in sequence. Specifically, it reconstructs the PLS latent variable features. After adding channel dimensions, the resulting size is The input tensor is fed into a one-dimensional convolutional layer. The one-dimensional convolutional layer is used to perform local combination mapping of latent variable features of different orders to enhance the expressive power of nonlinear associations between latent variables.

[0034] In one specific embodiment, the one-dimensional convolutional layer has 1 input channel, 16 output channels, a kernel size of 3, 1 edge padding, and an output size of [missing information]. The convolutional feature matrix.

[0035] The output of the one-dimensional convolutional layer serves as the input to the batch normalization layer. The batch normalization layer normalizes the feature channels of the convolutional layer output to reduce distribution differences between different batches of samples and different feature channels, thereby improving network training stability. The batch normalization layer does not change the feature matrix size; its output remains the same. And input it into the Gaussian error linear unit activation function layer.

[0036] The Gaussian error linear unit activation function layer is used to perform nonlinear mapping on the batch-normalized convolutional features, and its functional form is as follows: in, This is the cumulative distribution function of the standard normal distribution. This activation function is located after the batch normalization layer and before the bidirectional long short-term memory network layer, used to enhance the model's ability to express the nonlinear relationship between Raman spectral features and target concentration. The feature matrix after processing by the Gaussian error linear unit activation function remains... .

[0037] Subsequently, the feature matrix is ​​converted into The sequence features are then input into a bidirectional long short-term memory network layer. The sequence length is represented by 16, and the feature dimension corresponding to each sequence position is represented by 16. The bidirectional long short-term memory network layer includes a forward long short-term memory network and a backward long short-term memory network, which are used to extract positive and negative association information from the latent variable feature sequence, respectively.

[0038] In one specific embodiment, the bidirectional long short-term memory network layer adopts a single-layer bidirectional structure with an input dimension of 16, 64 hidden units in each direction, and an output dimension of 128. The final hidden states of the forward long short-term memory network and the backward long short-term memory network are concatenated to form the final reconstructed features. And input it into the Bayesian approximate neural network prediction unit.

[0039] The bidirectional long short-term memory network layer is an existing bidirectional LSTM network structure, which includes an input gate, a forget gate, an output gate, and a cell state update unit. The input gate controls the degree to which the current input features are written into the cell state, the forget gate controls the degree to which historical states are retained, the output gate controls the degree to which hidden states are output, and the cell state is used to transmit long-term dependency information in the sequence. In addition to the above-mentioned bidirectional long short-term memory network, unidirectional long short-term memory networks, gated recurrent unit networks, recurrent neural networks, or their equivalent structures that can achieve sequence feature association extraction can all be used as alternative structures in this invention.

[0040] The Bayesian approximate neural network prediction unit is used to receive the final reconstructed features. The system outputs the predicted concentrations of D-serine and L-serine in a single forward propagation. During the prediction phase, multiple Monte Carlo forward propagations are used to obtain multiple prediction results, and the final predicted concentration and prediction uncertainty are obtained based on the mean and standard deviation of these multiple prediction results. The Bayesian approximate neural network prediction unit comprises, in sequence, a first fully connected layer, a batch normalization layer, a Gaussian error linear unit activation function layer, a first random deactivation layer, a second fully connected layer, a Gaussian error linear unit activation function layer, a second random deactivation layer, an output layer, and a normalized output function layer.

[0041] In one specific embodiment, the final reconstructed features The dimension is 128. The first fully connected layer maps the 128-dimensional features to 64-dimensional features, used to compress and reorganize the comprehensive features output by the bidirectional long short-term memory network; the batch normalization layer is used to stabilize the distribution of the 64-dimensional features; the Gaussian error linear unit activation function layer is used to introduce nonlinear expression; and the first random deactivation layer is used to randomly mask some neuron connections. Subsequently, the second fully connected layer maps the 64-dimensional features to 32-dimensional features, which are then processed by the Gaussian error linear unit activation function layer and the second random deactivation layer before being input to the output layer. The output layer maps the 32-dimensional features to a 2-dimensional output, corresponding to the normalized predicted concentrations of D-serine and L-serine, respectively. In one specific embodiment, the normalized output function layer uses the Sigmoid function to restrict the prediction results to the range of 0 to 1; in other embodiments, a non-negative constraint function or truncation after linear output can also be used. Subsequently, inverse normalization is performed based on the concentration range in the training samples to obtain the actual concentration prediction value.

[0042] The Bayesian approximate neural network prediction unit is preferably a Bayesian approximate neural network based on Monte Carlo random deactivation. During the training phase, the random deactivation layer is enabled; during the prediction phase, the random deactivation layer remains enabled, and the batch normalization layer normalizes the data using the mean and variance saved during the training phase, allowing multiple prediction results for the same sample in multiple forward propagations. Each forward propagation is equivalent to sampling a sub-network prediction result from an approximate posterior distribution. The average of the multiple prediction results is used as the final predicted concentration, and the standard deviation of the multiple prediction results is used as the prediction uncertainty. It should be noted that the Bayesian approximate neural network is not limited to a Monte Carlo random deactivation structure; variational Bayesian approximate neural networks, deep ensemble neural networks, neural networks with probability output layers, or other equivalent network structures capable of simultaneously outputting the prediction mean and prediction uncertainty can also be used.

[0043] The model training phase includes sample preparation, training data construction, PLS feature extraction, neural network parameter updating, and model saving. First, standard samples with different combinations of D-serine and L-serine concentrations are prepared. Enhanced Raman spectra of each standard sample are acquired to form a training dataset with real concentration labels. Each training spectrum corresponds to a two-dimensional concentration label. in, Indicates the first The true concentration of D-serine in each training sample This indicates the actual concentration of L-serine. The label for training samples can be directly determined from the standard concentration used during sample preparation.

[0044] Secondly, the training spectra are sequentially subjected to SNV transformation, standardization, and PLS feature extraction to obtain the PLS latent variable features of the training samples. To improve the model's robustness to noise and slight baseline drift, data augmentation can be performed on the training spectra, including adding small-amplitude Gaussian noise and low-amplitude baseline drift. The augmented spectra retain the true concentration labels of the original samples and are converted into latent variable features using the same PLS model before participating in neural network training. Data augmentation is only applied to the training set samples; the validation set, test set, and samples to be predicted are not augmented.

[0045] The adjustable parameters in the one-dimensional convolutional neural network-LSTM feature reconstruction module include, but are not limited to: the number of output channels of the one-dimensional convolutional layer, the kernel size, the number of edge paddings, the number of hidden units in the long short-term memory neural network, the number of network layers, and whether a bidirectional structure is used. The adjustable parameters in the Bayesian approximate neural network include, but are not limited to: the number of fully connected layers, the number of neurons per layer, the activation function type, the random inactivation probability, and the number of Monte Carlo forward propagations. (The last sentence appears to be incomplete and possibly refers to the number of latent variables in PLS.) These are hyperparameters, determined through cross-validation before training the neural network weights; the weight parameters in convolutional layers, long short-term memory network layers, and fully connected layers are updated during training based on the loss value through backpropagation.

[0046] Before training, the true concentration labels are normalized to reduce the impact of different concentration ranges on the loss function. Preferably, the maximum concentrations of D-serine and L-serine in the training set are used as normalization factors to obtain the normalized labels: in, and These represent the maximum concentrations of D-serine and L-serine in the training set, respectively. The neural network output is the normalized prediction result: During training, the normalized prediction results output by the model are compared with the true normalized concentration labels, and the loss value is calculated based on the error between the two. Preferably, an improved dual-output weighted robust constraint loss function is used as the training loss function: in, For model training loss, For model parameters, The number of training samples. and The first Normalized true concentrations of D-type and L-type chiral target substances in each sample and These are the corresponding normalized predicted concentrations. and These are the weighting coefficients for D-type and L-type prediction errors, respectively, and γ is the weighting coefficient for the total concentration consistency constraint term. In a specific embodiment, and Set 1, and γ to 0.1. For model parameter regularization, The Huber loss function is defined as follows: in, This represents the error between the predicted value and the true label. In one specific embodiment, the threshold for the loss function is... We set it to 0.1. The Huber loss function approximates the mean squared error when the error is small and the absolute error when the error is large. Therefore, it can reduce the impact of abnormal spectra or abnormal labels on model training while ensuring regression accuracy.

[0047] In each training round, a batch of PLS ​​latent variable features are input into the model to obtain the normalized predicted concentrations of D-serine and L-serine. The loss value is calculated based on the predicted concentration and the true concentration label. Then, the weight parameters in the one-dimensional convolutional layer, long short-term memory network layer, and Bayesian approximate neural network are updated through backpropagation based on the loss value. After training is completed, the model parameters with the lowest loss value or the best validation performance are saved for subsequent prediction of test samples.

[0048] In one specific embodiment, the AdamW optimizer, i.e., an adaptive moment estimator optimizer with decoupled weight decay, is used during model training, with a learning rate of [missing information]. The batch size is 32, and the maximum number of training rounds is 600; the number of input channels in the one-dimensional convolutional layer is 1, the number of output channels is 16, the kernel size is 3, and the number of edge padding is 1; the long short-term memory neural network layer adopts a single-layer bidirectional structure with 64 hidden units; the fully connected layer structure of the Bayesian approximate neural network is 128→64→32→2, and the random deactivation probability is 0.3.

[0049] In the prediction phase, the spectrum to be tested is processed using the same SNV transformation, standardized parameters, and PLS model as in the training phase to obtain the PLS latent variable characteristics of the sample. Subsequently, the random deactivation layer in the Bayesian approximate neural network is kept enabled, and the batch normalization layer is normalized using the mean and variance saved in the training phase. The same spectrum to be tested is repeatedly subjected to T forward predictions, preferably T=100, to obtain multiple sets of prediction results. The prediction mean and standard deviation are as follows: in, As the final predicted concentrations of D-serine and L-serine, As a prediction uncertainty, the total concentration and the proportion of D-type can be calculated based on the predicted concentrations of D-serine and L-serine. Model performance is evaluated using metrics such as the coefficient of determination, root mean square error (RMSE), limit of detection (LOD), and limit of quantitation (LOQ). The coefficient of determination assesses the consistency between the model's predicted values ​​and the actual values, while the RMSE assesses the magnitude of the model's prediction error. Furthermore, the LOD and LOQ can be calculated based on the prediction results of a blank sample with an actual concentration of 0. Let the mean of the model's predictions for the blank sample be... The standard deviation is ,but: in, The detection limit is the lowest detectable concentration that can be distinguished from the blank response against a blank sample background. The limit of quantitation (LOQ) is the lowest concentration at which reliable quantitative analysis can be performed.

[0050] According to a sixth aspect of the present invention, a SERS-based chiral target detection method is provided, comprising: The chiral gold nanoparticles were mixed and incubated with the test sample containing the chiral target material in a liquid phase system to obtain a mixture; The mixture is applied to the surface of the enhanced Raman spectroscopy detection substrate, and surface-enhanced Raman spectroscopy detection is performed to achieve the detection of chiral target substances; The chiral gold nanoparticles have the same chiral configuration as the chiral target material.

[0051] In one embodiment, the chiral target substance comprises a chiral amino acid.

[0052] Furthermore, the chiral amino acid includes, but is not limited to, one or more combinations of cysteine, histidine, methionine, serine, threonine, tryptophan, and tyrosine.

[0053] Preferably, the chiral amino acid is serine, and more preferably D-serine.

[0054] In one embodiment, the mixed incubation time is 5 to 10 minutes.

[0055] In one embodiment, when performing the surface-enhanced Raman spectroscopy detection, the Raman detection power is 10~100 mW and the integration time is 1~5 s.

[0056] In one implementation scheme, the method specifically includes: The collected surface-enhanced Raman spectroscopy detection data is preprocessed, feature-selected, and concentration-predicted using a spectral data intelligent analysis module, thereby enabling quantitative analysis of the chiral target substance. The intelligent spectral data analysis module employs a deep Bayesian approximation neural network model, which is based on a 200–2000 cm⁻¹ depth. -1 Raman fingerprint region spectrum established.

[0057] According to a seventh aspect of the present invention, a SERS-based chiral target detection system is provided, comprising: The chiral gold nanoparticles have the same chiral configuration as the chiral target material; The enhanced Raman spectroscopy detection substrate; And, Raman spectrometer.

[0058] In one embodiment, the chiral target detection system further includes: The intelligent spectral data analysis module preprocesses, filters features, and predicts concentrations of the collected surface-enhanced Raman spectroscopy data, thereby enabling quantitative analysis of the chiral target substance.

[0059] According to an eighth aspect of the invention, the use of the SERS-based chiral target detection method or the SERS-based chiral target detection system in the detection of D-serine is provided.

[0060] According to a ninth aspect of the present invention, the application of the SERS-based chiral target detection method or the SERS-based chiral target detection system in rapid auxiliary screening for depression-related biomarkers is provided.

[0061] Compared with the prior art, the present invention has at least the following beneficial effects: (1) This invention utilizes the isochirality recognition capability of chiral gold nanoparticles to achieve preferential capture of chiral target substances, thereby improving chiral recognition selectivity. Furthermore, by coupling chiral gold nanoparticles with a glass fiber substrate modified with a three-dimensional network of silver nanoparticles to enhance Raman spectroscopy detection, a chiral surface-enhanced Raman detection platform with high sensitivity, high selectivity, and high signal amplification capability is constructed, which can form a stronger local electromagnetic field and a higher density hotspot, thus significantly enhancing the Raman signal. In addition, by introducing a deep Bayesian approximate neural network to intelligently analyze the full-spectral information in complex biological matrices, background interference is effectively reduced, and quantitative accuracy and robustness are improved.

[0062] (2) This invention exhibits good recognition ability for amino acids such as cysteine, histidine, methionine, serine, threonine, tryptophan, and tyrosine. The detection limit for D-serine in simulated urine and serum can reach 0.1 μM, and the recovery rate is mainly distributed between 90% and 110%. When the detection results were compared with those of liquid chromatography, the difference was less than 15%, which shows good prospects for clinical auxiliary screening. It can be used for rapid and non-invasive auxiliary screening of biomarkers related to depression and Alzheimer's disease, and provides a new technical path for the immediate detection of chiral molecules. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a schematic diagram illustrating the principle of a SERS-based chiral target detection method in one embodiment of the present invention. Figure 2 This is a TEM image of a gold octahedral seed crystal prepared in one embodiment of the present invention; Figure 3 This is a TEM image of L-shaped chiral gold nanoparticles prepared in one embodiment of the present invention. Figure 4 This is a TEM image of D-type chiral gold nanoparticles prepared in one embodiment of the present invention; Figure 5 The UV-Vis absorption spectra of gold octahedral seed crystals, L-configuration chiral gold nanoparticles, and D-configuration chiral gold nanoparticles prepared in one embodiment of the present invention are shown. Figure 6 The circular dichroism spectra of L-configuration chiral gold nanoparticles and D-configuration chiral gold nanoparticles prepared in one embodiment of the present invention are shown below. Figure 7 This is a photograph of a glass fiber membrane according to an embodiment of the present invention; Figure 8 This is an electron microscope image of a silver nanoparticle-deposited glass fiber substrate according to an embodiment of the present invention; Figure 9 These are Raman spectra of chiral serine detected by different methods in one embodiment of the present invention; Figure 10 This is a modeling flowchart of a spectral data intelligent analysis module according to an embodiment of the present invention; Figure 11This is a Raman peak selection diagram of a machine learning algorithm in one embodiment of the present invention; Figure 12 This is a linear regression analysis graph showing the relationship between the model-predicted concentration and the actual concentration in one embodiment of the present invention; Figure 13 This is a residual distribution diagram of a model prediction in one embodiment of the present invention; Figure 14 This is a histogram of the prediction error distribution of a model in one embodiment of the present invention; Figure 15 This is a diagram showing the relationship between a Bayesian neural network sub-model and an error in one embodiment of the present invention. Figure 16 This is an enhanced Raman spectroscopy pattern of D-serine in artificial urine in one embodiment of the present invention; Figure 17 This is an enhanced Raman spectroscopy pattern of D-serine in artificial serum in one embodiment of the present invention; Figure 18 This is a correlation analysis graph between the spiked concentration and the predicted concentration in one embodiment of the present invention; Figure 19 This is an error graph showing the detection results of this method and the GC-MS detection results in a serum sample in one embodiment of the present invention.

[0065] Figure 20 This is a significance analysis of D- / L-Ser concentration in serum samples in one embodiment of the present invention.

[0066] Figure 21 This is a bar chart showing the relative proportion of D- / L-Ser in a serum sample according to one embodiment of the present invention.

[0067] Figure 22 This is a heatmap showing the relative proportion of D- / L-Ser in a serum sample according to one embodiment of the present invention. Detailed Implementation

[0068] To address the shortcomings of existing technologies, this application provides a chiral target detection method based on chiral gold nanoparticles and a silver-modified glass fiber surface-enhanced Raman spectroscopy substrate. The method utilizes the adsorption effect of chiral gold nanoparticles on chiral target substances (such as chiral amino acid molecules) with the same optical rotation. Combined with the silver-modified glass fiber surface-enhanced Raman spectroscopy substrate for Raman spectroscopy detection of chiral target substances, this significantly improves the detection signal intensity and sensitivity. Furthermore, an intelligent spectral data analysis module based on a machine learning network architecture processes and analyzes the acquired Raman spectral detection signals, further effectively improving the accuracy of quantitative analysis of chiral target substances under complex conditions.

[0069] Specifically, this application provides a SERS-based method for chiral target detection. See [reference needed]. Figure 1 As shown, the method includes: First, chiral gold nanoparticles are induced using a chiral dipeptide template, and a substrate for enhanced Raman spectroscopy detection is prepared via a silver mirror reaction. A deep Bayesian approximate neural network module for full-spectrum modeling and analysis is then constructed, thus forming an intelligent spectral data analysis module.

[0070] Secondly, during the detection process, chiral gold nanoparticles are first mixed with the sample containing chiral target material and incubated. This allows the target chiral material, which has the same chirality as the chiral gold nanoparticles, to be preferentially adsorbed onto the surface of the sample. The complex of the two is then transferred to an enhanced Raman spectroscopy detection substrate. A high-intensity electromagnetic hotspot is formed by gold-silver bimetallic plasmon coupling to collect the enhanced Raman signal. Finally, the obtained full fingerprint spectrum is imported into a machine learning network model to achieve high-precision concentration prediction in complex samples.

[0071] The technical solution of the present invention will be further described in detail below with reference to several preferred embodiments and accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. It should be noted that the following embodiments are intended to facilitate the understanding of the present invention, and do not constitute any limitation thereof. 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. Experimental methods in the following embodiments that do not specify specific conditions are generally performed under conventional conditions or according to the conditions recommended by the manufacturer. Temperatures not mentioned are assumed to be room temperature conditions (e.g., within the range of 15-35°C, preferably around 25°C). The experimental conditions and set parameters therein should not be regarded as limitations on the basic technical solution of the present invention.

[0072] Example 1: Preparation of gold octahedral seed crystals Prepare 20 mL of an aqueous solution containing 0.125 mM chloroauric acid and 10 mM CTAB. Add 100 μL of ascorbic acid solution (100 mM) to this solution. After the initial pale yellow color disappears, quickly inject 100 μL of sodium hydroxide aqueous solution (100 mM) into the mixture to initiate particle formation. Under gentle shaking, the solution color gradually changes from pale pink to purplish-red within 20 min. Throughout the reaction, maintain the temperature in a water bath within the range of 25–60 °C. The resulting gold octahedral nanoparticles are collected by centrifugation at 7000 r / min and repeatedly dispersed and washed three times with 1 mL of CTAB solution (1 mM) to remove excess reactants, obtaining a gold octahedral seed dispersion. The morphology and size of the gold octahedral seeds are as follows: Figure 2 As shown.

[0073] Example 2: Preparation of chiral gold nanoparticles The growth solution was prepared by mixing 1 mL of CTAB solution (10 mM) with 4 mL of deionized water. Then, 80 μL of chloroauric acid solution (25 mM) was added to this growth solution. After incubating the solution for 10 min, 500 μL of L-ascorbic acid solution (40 mM) was rapidly injected to remove Au. 3+ Restore to Au + Subsequently, 5 μL of chiral dipeptide solution (containing 4 mM cysteine-leucine dipeptide) and 20 μL of the aforementioned gold octahedral seed dispersion were added. The mixture was thoroughly shaken and reacted at room temperature for 1 h. Finally, the synthesized chiral gold nanoparticles were collected by centrifugation at 7000 r / min and washed three times with 1 mL CTAB solution (1 mM) to remove unreacted substances, obtaining a chiral gold nanoparticle dispersion. The morphology, UV-Vis absorption spectrum, and circular dichroism spectroscopy of the chiral gold nanoparticles are shown below. Figure 3 , 4 As shown in Figures 5 and 6.

[0074] Example 3: Preparation of substrate for enhanced Raman spectroscopy detection The glass reaction vessel was placed in a dilute sodium hydroxide solution and pretreated by boiling under stirring to remove impurities, followed by thorough rinsing with deionized water. 5 mL of silver nitrate solution (2%) was added to the cleaned vessel. Subsequently, ammonia solution (2%) was added dropwise under continuous stirring until the initially formed brown precipitate just dissolved, yielding a clear silver ammonia solution. Next, glucose was dissolved in a 10% (v / v) aqueous ethanol solution to form a reducing solution (glucose concentration 200 mM).

[0075] Then, the glass fiber membrane (such as...) Figure 7 The glass fiber membrane (as shown) was immersed in freshly prepared silver ammonia solution for 30 min to ensure sufficient adsorption of silver ammonia ions. Subsequently, the membrane was rinsed with distilled water to remove loosely bound silver ions. Next, the pretreated glass fiber was immersed in the aforementioned reduction solution and incubated at 45°C for 30 min to promote the reduction of silver ions. After the reaction was complete, the resulting silver-plated glass fiber was washed sequentially with double-distilled water and anhydrous ethanol, thus obtaining the substrate for enhanced Raman spectroscopy detection. Finally, the substrate was dried in a 37°C oven and stored for later use. An electron micrograph of the substrate is shown below. Figure 8 As shown.

[0076] Example 4: Amino Acid Chiral Recognition and Detection The D-type and L-type Ser stock solutions (0.1 M) were prepared separately in deionized water for later use. In a typical detection system, the aforementioned chiral gold nanoparticle dispersion was mixed with a serine solution and incubated at room temperature for 10 min to promote the full adsorption and interaction of serine molecules with the surface of the chiral gold nanoparticles. Subsequently, the mixture was dropped onto the surface of the aforementioned silver-modified glass fiber reinforced Raman detection substrate for Raman spectroscopy acquisition.

[0077] Figure 9 The Raman spectral responses of D-serine in different detection systems are shown. Figure 9 It is evident that the Raman signal is weak when using a silver-modified glass fiber substrate alone to detect D-serine; however, the characteristic Raman peaks of D-serine are significantly enhanced upon the introduction of chiral gold nanoparticles. Specifically, the Raman signal of D-serine is strongest when the silver-modified glass fiber substrate and chiral D-type gold nanoparticles interact together, with multiple characteristic peaks showing significant enhancement; while the Raman enhancement effect of D-serine is relatively weak when using chiral L-type gold nanoparticles. These results indicate a stronger chiral matching effect between chiral D-type gold nanoparticles and D-serine, which can improve the enrichment of D-serine on the enhanced substrate surface and the Raman response intensity. In contrast, the Raman enhancement effect is reduced in the chiral mismatch system, suggesting that the constructed chiral gold nanoparticle-silver-modified glass fiber composite detection system can achieve differentiated recognition and enhanced detection of serine enantiomers.

[0078] Example 5: Machine Learning Model Prediction Method Surface-enhanced Raman spectra of the test samples were acquired to obtain the original Raman spectral vectors. To achieve quantitative analysis of complex Raman spectra, PyCharm was used for data processing. In this experiment, D-Ser and L-Ser concentration levels were set to 0, 0.1, 0.5, 1, 5, and 10 μM, respectively, resulting in 36 concentration combinations through full factorial analysis. After removing the blank group (where both D-Ser and L-Ser were 0 μM), 35 effective concentration combinations were obtained. Ten spectra were repeatedly acquired for each concentration combination and detected on both chiral-enhanced substrates, resulting in a total of 35 × 10 × 2 = 700 effective Raman spectra. Model training was completed using these 700 independent spectral samples. Figure 10 The model shown (its specific structure and function have been described above) has a prediction coefficient of determination R for D-serine. 2 The mean square error (RMSE) was 0.9925, the root mean square error (RMSE) was 0.1948, and the limit of detection (LOD) reached 0.1 μM. The results are as follows: Figures 11-14 As shown.

[0079] Example 6: Application Validation in Complex Samples Artificial urine and artificial serum were used as complex matrix models (both purchased from Sangon Biotech Co., Ltd.) to conduct spiked recovery experiments on the chiral target detection method of this invention. During detection, the sample matrix was first diluted 10-fold, then 40 μL of L-type chiral gold nanoparticles and D-serine (final concentrations of 20, 50, 100, 250, 500, 800, 1000, 1200, 1500, 2000, and 15000 μM) were added. After mixing and incubation at room temperature for 10 minutes, the mixture was dropped onto the surface of the enhanced Raman spectroscopy detection substrate for detection. The specific steps were the same as in Example 4. The detection results are as follows: Figures 15-18 As shown in the figure. The detection results show that the model-predicted concentration and the actual concentration are highly consistent at different spiking concentrations, and the recovery rate is mainly concentrated between 90% and 110%, indicating that the chiral target detection method of the present invention has good adaptability to complex matrices and practical application potential.

[0080] Example 7: Actual Sample Detection An Alzheimer's disease mouse model was established using previously reported methods with slight modifications. In short, to induce neuroinflammation, mice were intraperitoneally injected daily for 7 consecutive days with lipopolysaccharide (1 mg / kg); control mice were injected with an equal volume of sterile saline. Continuous gavage administration was performed during lipopolysaccharide administration to ensure sustained drug exposure. Behavioral assessments were conducted at week 8 following 7 days of lipopolysaccharide-induced neuroinflammation. Serum extraction was performed at the end of the process.

[0081] The detection steps are the same as in Example 6, and the results can be found in [reference needed]. Figures 19-22 As shown, the results indicate that the serum D-serine content in Alzheimer's disease mice was significantly increased, but the overall serine level remained essentially unchanged. This result showed no significant difference from the liquid chromatography results, demonstrating that the chiral target detection method of this invention has excellent reliability and stability.

[0082] Based on the above embodiments, it can be seen that the present invention provides a chiral target detection method that combines high selectivity, high sensitivity, high repeatability and good adaptability to complex matrices. It can not only achieve the priority identification of target chiral amino acids such as chiral D-serine, but also complete rapid quantitative analysis in complex samples by combining intelligent algorithms. Thus, it can well meet the application needs of auxiliary screening and immediate detection of biomarkers related to depression and Alzheimer's disease.

[0083] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0084] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A chiral target detection method based on SERS, characterized in that, include: Chiral gold nanoparticles were mixed and incubated with a sample containing a chiral target substance in a liquid phase system to obtain a mixture; The mixture is applied to the surface of a surface-enhanced Raman spectroscopy (SERS) substrate, and surface-enhanced Raman spectroscopy is performed to detect chiral target substances. The chiral gold nanoparticles have the same chiral configuration as the chiral target material; the enhanced Raman spectroscopy detection substrate includes a glass fiber membrane and a three-dimensional porous network structure bonded to the glass fiber membrane, wherein the three-dimensional porous network structure is formed by the stacking of silver nanoparticles.

2. The method according to claim 1, characterized in that, include: A gold octahedral seed crystal is provided, wherein the surface of the gold octahedral seed crystal is entirely composed of (111) crystal planes; Using a chiral dipeptide as a chiral template, gold atoms are induced to be deposited in a directional manner on the surface of the gold octahedral seed crystal, thereby forming the chiral gold nanoparticles.

3. The method according to claim 2, characterized in that, include: Under conditions of 25~60℃, an alkali is added to an aqueous solution containing 0.1~0.5 mM chloroauric acid, 10~100 mM ascorbic acid and 1~10 mM MCTAB to react and obtain the gold octahedral seed crystal, wherein the molar ratio of chloroauric acid to alkali is 0.5~2.

5. And / or, at room temperature, ascorbic acid is added to an aqueous solution containing 0.1–0.5 mM chloroauric acid and 1–10 mM CTAB to carry out the reaction, and the molar ratio of chloroauric acid to ascorbic acid is 0.5–2.5, thereby obtaining a product containing Au. + The growth solution is prepared by adding a chiral dipeptide and the gold octahedral seed crystals to the growth solution and mixing them evenly, so that the concentration of the chiral dipeptide is 0.5~4 mM and the concentration of the gold octahedral seed crystals is 0.2~1 OD, and then reacting for more than 1 hour to obtain the chiral gold nanoparticles. And / or, the chiral dipeptide includes a cysteine-leucine dipeptide.

4. The method according to claim 1, characterized in that, include: Silver nanoparticles are deposited on a glass fiber membrane via a silver mirror reaction to form the enhanced Raman spectroscopy detection substrate.

5. The method according to claim 4, characterized in that, Specifically, it includes: The glass fiber membrane is thoroughly immersed in silver ammonia solution, then removed and rinsed with deionized water. After immersion in a reducing solution, it is incubated at 25-60°C for 10-60 min to obtain the enhanced Raman spectroscopy detection substrate. The concentration of the silver ammonia solution, calculated as AgNO3, is 20-200 mM. The reducing solution is an aqueous solution containing 50-200 mM glucose.

6. The method according to claim 1, characterized in that, Specifically, it includes: A quantitative prediction model for chiral target substances was used to simultaneously predict the concentrations of D-type and L-type chiral target substances based on surface-enhanced Raman spectroscopy data. The quantitative prediction model for chiral target substances includes a latent variable extraction unit, a one-dimensional convolutional-bidirectional long short-term memory feature reconstruction unit, and a Bayesian approximate neural network prediction unit. The latent variable extraction unit is used to extract low-dimensional latent variable features from the surface-enhanced Raman spectroscopy data using partial least squares method, and inputs these features into the one-dimensional convolutional-bidirectional long short-term memory feature reconstruction unit. The one-dimensional convolutional-bidirectional long short-term memory feature reconstruction unit includes a one-dimensional convolutional layer, a batch normalization layer, a Gaussian error linear unit activation layer, and a bidirectional long short-term memory neural network layer. The one-dimensional convolutional layer is used to extract local combination features between adjacent latent variables, and the bidirectional long short-term memory neural network layer is used to extract forward and backward dependencies in the latent variable sequence, thereby obtaining a deep spectral characterization input into the Bayesian approximate neural network prediction unit. The Bayesian approximate neural network prediction unit includes a first fully connected layer, a batch normalization layer, a Gaussian error linear unit activation layer, a first random deactivation layer, a second fully connected layer, a Gaussian error linear unit activation layer, a second random deactivation layer, and an output layer connected in sequence. The output layer includes two output nodes, corresponding to the concentrations of D-type chiral target substances and L-type chiral target substances, respectively; The quantitative prediction model for chiral target substances is implemented using supervised training, specifically including: The optimal number of latent variables was determined by cross-validation, and the model parameters of the one-dimensional convolutional-bidirectional long short-term memory feature reconstruction unit and the Bayesian approximate neural network prediction unit were optimized by using a dual-output weighted robust constraint loss function. The dual-output weighted robust constraint loss function is expressed as follows: in, For model training loss, For model parameters, The number of training samples. and The first Normalized true concentrations of D-type and L-type chiral target substances in each sample and These are the corresponding normalized predicted concentrations. and These are the weighting coefficients for the prediction errors of type D and type L, respectively. The weighting coefficients for the total concentration consistency constraint term. For model parameter regularization, Here is the Huber robust error function, defined as follows: in, For prediction error, This is the error threshold.

7. The method according to claim 1, characterized in that: The chiral target substance includes chiral amino acids, which include one or more combinations of cysteine, histidine, methionine, serine, threonine, tryptophan, and tyrosine. And / or, the incubation time for the mixture is 5 to 10 minutes.

8. A chiral target detection system based on SERS, characterized in that, include: Chiral gold nanoparticles, wherein the chiral gold nanoparticles have the same chiral configuration as the chiral target material; An enhanced Raman spectroscopy detection substrate includes a glass fiber membrane and a three-dimensional porous network structure bonded to the glass fiber membrane, wherein the three-dimensional porous network structure is formed by the stacking of silver nanoparticles; And, Raman spectrometer.

9. The SERS-based chiral target detection system according to claim 8, characterized in that, Also includes: The intelligent spectral data analysis module preprocesses, filters features, and predicts concentrations of the collected surface-enhanced Raman spectroscopy data, thereby enabling quantitative analysis of the chiral target substance.

10. The use of the SERS-based chiral target detection method according to any one of claims 1-7 or the SERS-based chiral target detection system according to any one of claims 8-9 in the detection of D-serine.

Images