Pengju Yin

The surface-enhanced Raman spectroscopy (SERS)-based electronic tongue (E-tongue) represents a noninvasive and label-free taste-mimicking technology. This method employs an SERS sensing array as the core for signal amplification and detection, integrated with spectral signal acquisition and machine learning (ML) techniques, to achieve specific identification, qualitative analysis, and quantitative determination of multiple components in complex liquid matrices. However, extracting key features and information from complex, multi-component mixture spectra and selecting appropriate analytical methods based on data characteristics and specific tasks remain significant challenges in the field. This paper reviews the technical principles and research progress of SERS-based E-tongues, focusing on the core issue of "How laboratory researchers can scientifically and efficiently select ML methods after obtaining raw SERS data from sensor arrays, based on the data characteristics and analysis goals." The sections on preprocessing and feature engineering systematically summarize mainstream methods and evaluate their applicable scenarios. The qualitative analysis and quantitative modeling section also provides clear guidelines for selecting algorithms based on typical spectral characteristics and specific task requirements. The entire discussion addresses the practical challenges throughout the data processing pipeline, aiming to provide clear and practical methodological references for the majority of laboratory personnel engaged in SERS-based E-tongue research. Finally, it highlights the latest applications, summarizes current bottlenecks, and outlines future directions.

Quantitative protein analysis by surface-enhanced Raman spectroscopy (SERS) remains challenging due to weak and heterogeneous protein adsorption on plasmonic surfaces. Here, we introduce a superspectra-guided SERS framework that leverages chemically distinct interaction environments to enhance quantitative performance. Silver nanorod (AgNR) substrates were functionalized with cysteamine (CM), cysteine (CN), and 6-mercapto-1-hexanol (MCH), together with unmodified (B) AgNRs, to create surfaces that probe complementary aspects of protein–surface interactions through charge- and chemistry-dependent binding. Using bovine serum albumin (BSA) as a model protein, we systematically constructed superspectra by concatenating SERS signals from all single-, pairwise-, triple-, and four-surface combinations and evaluated their performance using support vector regression (SVR) and random forest regression (RFR). Our results reveal that superspectra must be constructed selectively: single-substrate spectra lack sufficient chemical diversity, and superspectra incorporating all four surfaces often degrade accuracy due to noninformative or conflicting features, particularly those introduced by MCH. In contrast, superspectra derived from complementary surface chemistries, especially the CM&CN pair or the B&CM&CN triplet, yield markedly improved quantitative predictions. RFR consistently outperformed SVR, demonstrating superior robustness for integrating chemically heterogeneous spectral inputs. This work establishes, for the first time, design principles for constructing effective superspectra for protein SERS and highlights the importance of analyte–surface interaction complementarity in enabling accurate, scalable protein quantification.

With the rapid development of science and technology, artificial intelligence (AI) and machine learning (ML) have become the forefront of innovation. Electronic noses and tongues, which mimic human smell and taste perception, have great potential in fields such as environmental monitoring, food safety, medical diagnosis and quality control. However, the traditional electronic nose and tongue systems still face challenges due to the complexity of marker morphology and the limitations of sensors. Machine learning technology has also proven to be a valuable tool for improving traditional electronic nose and tongue technology. In this review, we introduce the principle and design of machine learning combined with electronic nose and tongue technology, analyze some recently published articles on machine learning-assisted electronic nose and tongue for marker detection, and review the practical application of machine learning technology in the field of electronic nose and tongue. It is believed that through continuous exploration and innovation, machine learning will promote the application of electronic nose and electronic tongue, realize intelligent monitoring and control, and provide help for human life and health.

Neuromyelitis optica spectrum disorder (NMOSD) diagnosis currently relies on the cell-based assay (CBA) as the gold standard for detecting aquaporin-4 (AQP4) antibodies. However, manual interpretation of AQP4 antibody immunofluorescence images is subjective, time-consuming, and prone to variability. These images often exhibit challenges such as small, densely distributed positive signals, weak fluorescence intensity, and severe background interference from non-specific binding or artifacts.To address these issues, we propose an improved YOLO-based model, termed YOLO-SCS, built upon YOLOv11, for accurate detection of positive signals in AQP4 immunofluorescence images. First, a high-quality dataset of AQP4 immunofluorescence images was constructed with precise annotations. A Gaussian Peak-optimized Channel Adjustment (GPCA) algorithm was designed for preprocessing to enhance weak fluorescence contrast while effectively suppressing background noise. On the YOLOv11 architecture, we incorporated the SPD-Conv module in the downsampling stages to preserve fine-grained spatial features critical for small targets. Additionally, a shape-aware Normalized Wasserstein Distance (NWD) loss function was adopted to improve bounding box regression accuracy for densely overlapping and irregularly shaped positive regions. Experimental evaluations on a clinically representative dataset (positive-to-negative ratio of 1:9) demonstrated superior performance: the proposed YOLO-SCS achieved a positive detection rate of 98.4%, negative specificity of 100%, mAP@0.5 of 0.6971, and mAP@0.5:0.95 of 0.4256. These metrics significantly outperform the baseline YOLOv11 and other state-of-the-art improved models. The model reliably distinguishes specific positive fluorescence signals from background impurities, providing robust automated interpretation support for AQP4 immunofluorescence in NMOSD clinical diagnosis and potentially reducing diagnostic turnaround time and inter-observer variability.

The application of machine learning to Raman spectral analysis is limited by the scarcity of labeled data, particularly in single-cell studies where large datasets are difficult to obtain. Under such small-sample conditions, the reliability of different data augmentation strategies remains unclear. This study systematically evaluates four data augmentation methods—localized blurring, Gaussian noise addition, random amplitude scaling, and generative adversarial network (GAN)–based synthesis—for small-scale Raman spectral classification of breast cells, focusing on a training set containing 10 samples per class. Distributional similarity between original and augmented data was assessed using Fréchet Inception Distance and t-distributed stochastic neighbor embedding, and classification performance was evaluated using a one-dimensional ResNet model. The results show that augmentation effectiveness depends on both the augmentation strategy and the number of synthetic samples. Gaussian noise augmentation achieved the highest distributional similarity and improved classification accuracy from 92.45% to 95.35%, while localized blurring also yielded consistent improvements, with accuracies exceeding 94.30%. GAN-based augmentation enhanced performance at suitable augmentation sizes, reaching peak accuracies above 94.10%, but showed greater sensitivity to parameter selection. In contrast, random amplitude scaling provided little improvement across most settings. Additional parameterization experiments were conducted by changing the original training set to investigate extremely data-scarce scenarios, including 1 or 2 spectra per class. In these cases, data augmentation alleviated severe overfitting, with Gaussian and blurring methods providing the most stable gains, whereas GAN-based augmentation showed variable effectiveness and scaling remained ineffective. These results offer practical, quantitative guidance for selecting appropriate data augmentation strategies in small-scale Raman spectral analysis.

Raman spectroscopy has become an important tool for biomedical analysis due to its ability to provide label-free, non-destructive molecular fingerprints of biological samples. However, existing deep learning approaches for classifying biological Raman spectra often focus on specific datasets and lack generalizability and interpretability. In this study, BioRamanNet is presented, an interpretable and generalizable deep learning framework designed for classifying a wide range of biological Raman spectra. The model integrates adaptive one-dimensional convolutional layers and squeeze-and-excitation (SE) blocks within a residual network architecture to enhance feature extraction. BioRamanNet was evaluated using four representative Raman spectral datasets—breast cells, extracellular vesicles and particles (EVPs), viruses, and bacteria—achieving classification accuracies of 99.5%, 100%, 99.8%, and 85.3%, respectively. To improve model interpretability, a perturbation-based analysis using Voigt noise was introduced to identify key wavenumber regions influencing classification. These regions were found to correspond closely with known Raman biomarkers, validating their biological significance. The results of this work demonstrate that BioRamanNet is a powerful and interpretable tool for analyzing diverse biological Raman spectra and holds promise for advancing machine learning-assisted biomedical diagnostics.

Raman spectroscopy has emerged as a pivotal technology in modern scientific research and industrial applications, offering nondestructive, high-resolution analysis with robust molecular fingerprinting capabilities. The extraction of Raman spectral features is a critical step in spectral data analysis, directly influencing sample identification, classification, and quantitative outcomes. However, integrating important data features from machine learning models with context-specific biosignatures to derive meaningful insights into spectral analysis remains a significant challenge. Herein, the Raman Peak Feature Matching (RPFM) method is proposed, which matches protein peak features with salient breast cell data features extracted from the machine learning models. Feature augmentation is subsequently applied to the matching-retained breast cell features, thereby enhancing spectral analysis capabilities. The RPFM method is applied to breast cell spectra for feature augmentation with a reclassification accuracy of 97.12% using a linear support vector machine model, achieving an 8.34% improvement over the model's performance without feature augmentation. The RPFM method has also been successfully implemented in generalized linear logistic regression and tree-based eXtreme gradient boosting, demonstrating its versatility across diverse machine learning algorithms. The RPFM method leverages data-driven machine learning models while compensating for the inability to take into account specific specialized background knowledge. This methodology significantly advances the accuracy and efficacy of spectral analysis in biological and medical applications, offering a novel framework for machine learning algorithms to perform augmented Raman spectral analysis.

Surface-enhanced Raman scattering (SERS) has emerged as a potent spectroscopic technique for the detection of single cells. However, it is difficult to achieve label-free detection at the single-cell level in dynamic liquids because nanoprobe aggregation in biological fluids and the low combination of nanoprobes and cells reduce the sensitivity of SERS detection. Herein, a dynamic liquid integrated single-cell SERS (DLISC-SERS) platform is developed for the label-free detection of single cancer cells. DLISC-SERS consists of three components, including a twisted mixing microfluidic chip to achieve an efficient combination of nanoprobes and cells, a commercial coaxial needle to accomplish 3D dynamic liquid focusing by annular sheath flow, and a quartz capillary to offer a SERS detection area with low noise. The mixing intensity of the twisted mixing microfluidic chip is almost 3.67-fold higher than that of straight mixing. The multifunctionally modified nanoprobe, Ag NSs@PEG@3COOH, can be stably dispersed in biological fluids for at least 30 min. The segment weighting similarity-based KNN model can classify single-cell spectra with sensitivity, specificity, and accuracy up to 100, 99.4, and 99.5%, respectively. The accuracy of the model for three-way classification is 95.2%. The DLISC-SERS platform is a powerful tool for detecting cancer cells at the single-cell level.