Local classification: Locally weighted–partial least squares-discriminant analysis (LW–PLS-DA)

Increased interpretation of deep learning models using hierarchical cluster-based modelling

Linear prediction models based on data with large inhomogeneity or abrupt non-linearities often perform poorly because relationships between groups in the data dominate the model. Given that the data is locally linear, this can be overcome by splitting the data into smaller clusters and creating a local model within each cluster. In this study, the previously published Hierarchical Cluster-based Partial Least Squares Regression (HC-PLSR) procedure was extended to deep learning, in order to increase the interpretability of the deep learning models through local modelling. Hierarchical Cluster-based Convolutional Neural Networks (HC-CNNs), Hierarchical Cluster-based Recurrent Neural Networks (HC-RNNs) and Hierarchical Cluster-based Support Vector Regression models (HC-SVRs) were implemented and tested on spectroscopic data consisting of Fourier Transform Infrared (FT-IR) measurements of raw material dry films, for prediction of average molecular weight during hydrolysis and a simulated data set constructed to contain three clusters of observations with different non-linear relationships between the independent variables and the response. HC-CNN, HC-RNN and HC-SVR outperformed HC-PLSR for the simulated data set, showing the disadvantage of PLSR for highly non-linear data, but for the FT-IR data set there was little to gain in prediction ability from using more complex models than HC-PLSR. Local modelling can ease the interpretation of deep learning models through highlighting differences in feature importance between different regions of the input or output space. Our results showed clear differences between the feature importance for the various local models, which demonstrate the advantages of a local modelling approach with regards to interpretation of deep learning models.

SERS Signature of SARS-CoV-2 in Saliva and Nasopharyngeal Swabs: Towards Perspective COVID-19 Point-of-Care Diagnostics

In this study, the intrinsic surface-enhanced Raman spectroscopy (SERS)-based approach coupled with chemometric analysis was adopted to establish the biochemical fingerprint of SARS-CoV-2 infected human fluids: saliva and nasopharyngeal swabs. The numerical methods, partial least squares discriminant analysis (PLS-DA) and support vector machine classification (SVMC), facilitated the spectroscopic identification of the viral-specific molecules, molecular changes, and distinct physiological signatures of pathetically altered fluids. Next, we developed the reliable classification model for fast identification and differentiation of negative CoV(-) and positive CoV(+) groups. The PLS-DA calibration model was described by a great statistical value-RMSEC and RMSECV below 0.3 and R²_cal at the level of ~0.7 for both type of body fluids. The calculated diagnostic parameters for SVMC and PLS-DA at the stage of preparation of calibration model and classification of external samples simulating real diagnostic conditions evinced high accuracy, sensitivity, and specificity for saliva specimens. Here, we outlined the significant role of neopterin as the biomarker in the prediction of COVID-19 infection from nasopharyngeal swab. We also observed the increased content of nucleic acids of DNA/RNA and proteins such as ferritin as well as specific immunoglobulins. The developed SERS for SARS-CoV-2 approach allows: (i) fast, simple and non-invasive collection of analyzed specimens; (ii) fast response with the time of analysis below 15 min, and (iii) sensitive and reliable SERS-based screening of COVID-19 disease.

Discrimination of leukemias and non-leukemic cancers in blood serum samples of children and adolescents using a Raman spectral model

This study aimed to identify the differences presented in the Raman spectrum of blood serum from normal subjects compared to leukemic and non-leukemic subjects and the differences between the leukemics and non-leukemics, correlating the spectral differences with the biomolecules. Serum samples from children and adolescents were subjected to Raman spectroscopy (830 nm, laser power 350 mW; n = 566 spectra, being 72 controls, 269 leukemics, and 225 non-leukemics). Exploratory analysis based on principal component analysis (PCA) of the serum sample's spectra was performed. Classification models based on partial least squares discriminant analysis (PLS-DA) were developed to classify the spectra into normal, leukemic, and non-leukemic, as well as to discriminate spectra of leukemic from non-leukemic. The exploratory analysis showed principal components with peaks related to amino acids, proteins, lipids, and carotenoids. The spectral differences between normal, leukemic, and non-leukemic showed features assigned to proteins (serum features), amino acids, and carotenoids. The PLS-DA model classified the spectra of the normal group versus leukemic and non-leukemic groups with accuracy of 66%, sensitivity of 99%, and specificity of 57%. The PLS-DA discriminated the spectra of the leukemic and non-leukemic groups with accuracy of 67%, sensitivity of 72%, and specificity of 60%. The study showed that Raman spectroscopy is a technique that may be used for the biochemical differentiation of leukemias and other types of cancer in serum samples of children and adolescents. Nevertheless, building an extensive data library of Raman spectra from serum samples of controls, leukemics, and non-leukemics of different age groups is necessary to understand the findings better.

Chemometrics for Raman Spectroscopy Harmonization

Raman spectroscopy is used in a wide variety of fields, and in a plethora of different configurations. Raman spectra of simple analytes can often be analyzed using univariate approaches and interpreted in a straightforward manner. For more complex spetral data such as time series or line profiles (1D), Raman maps (2D), or even volumes (3D), multivariate data analysis (MVDA) becomes a requirement. Even though there are some existing standards for creation, implementation, and validation of methods and models employed in industry and academics, further research and development in the field must contribute to their improvement. This review will cover, in broad terms, existing techniques as well as new developments for MVDA for Raman spectroscopic data, and in particular the use associated with instrumentation and data calibration. Chemometric models are often generated via fusion of analytical data from different sources, which enhances model discrimination and prediction abilities as compared to models derived from a single data source. For Raman spectroscopy, raw or unprocessed data is rarely ever used. Instead, spectra are usually corrected and manipulated,¹ often by case-specific rather than universal methods. Calibration models can be used to characterize qualitatively and/or quantitatively samples measured with the same instrumentation that was used to create the model. However, regular validation is required to ensure that aging or incorrect maintenance of the instrument does not alter the model's predictions, particularly when applied in regulated fields such as pharmaceuticals. Furthermore, a model transfer may be required for different reasons, such as replacement or significant repair of the instrumentation. Modeling can also be used to consistently harmonize Raman spectroscopic data across several instrumental designs, accounting for variations in the resulting spectrum induced by different components. Data for Raman harmonization models should be processed in a protocolled manner, and the original data accessible to allow for model reconstruction or transfer when new data is added. Important processing steps will be the calibration of the spectral axes and instrument dependent effects, such as spectral resolution. In addition, data fusion and model transfer are essential for allowing new instrumentation to build on existing models to harmonize their own data. Ideally, an open access database would be created and maintained, for the purpose of allowing for continued harmonization of new Raman instruments using an outlined and accepted protocol.

Quality Prediction Model of KICA-JITL-LWPLS Based on Wavelet Kernel Function

To obtain quality variables that cannot be measured in real time during the production process but reflect information on the quality of the final product, the batch production process has the characteristics of a strong time-varying nature, non-Gaussian data distribution and high nonlinearity. A locally weighted partial least squares regression quality prediction model (KICA-JITL-LWPLS), based on wavelet kernel function independent meta-analysis with immediate learning, is proposed. The model first measures the similarity between the normalized input data and the historical data and assigns the input data to the group of historical data with high similarity to it, based on the posterior probability of the Bayesian classifier; subsequently, wavelet kernel functions are selected and kernel learning methods are introduced into the independent meta-analysis algorithm. An independent meta-analysis, based on the wavelet kernel function, is performed on the classified input data to obtain probabilistically significant independent sets of variables. Finally, a real-time learning-based LWPLS regression analysis is performed on this variable set to construct a local prediction model for the current sample by calculating the similarity between the local input data. The local predictions from the PLS output are fused with the posterior probability output from the Bayesian classifier to produce the final prediction. The method was used to predict the product concentration and bacteriophage concentration during penicillin fermentation through a simulation platform. The prediction results were basically consistent with the real values, proving that the proposed KICA-JITL-LWPLS quality prediction model, based on wavelet kernel functions, has reliable prediction results.

Feature selection of infrared spectra analysis with convolutional neural network

Data-driven deep learning analysis, especially for convolution neural network (CNN), has been developed and successfully applied in many domains. CNN is regarded as a black box, and the main drawback is the lack of interpretation. In this study, an interpretable CNN model was presented for infrared data analysis. An ascending stepwise linear regression (ASLR)-based approach was leveraged to extract the informative neurons in the flatten layer from the trained model. The characteristic of CNN network was employed to visualize the active variables according to the extracted neurons. Partial least squares (PLS) model was presented for comparison on the performance of extracted features and model interpretation. The CNN models yielded accuracies with extracted features of 93.27%, 97.50% and 96.65% for Tablet, meat, and juice datasets on the test set, while the PLS-DA models obtained accuracies with latent variables (LVs) of 95.19%, 95.50% and 98.17%. Both the CNN and PLS models demonstrated the stable patterns on active variables. The repeatability of CNN model and proposed strategies were verified by conducting the Monte-Carlo cross-validation.

Chemometric Strategies for Spectroscopy-Based Food Authentication

In the last decades, spectroscopic techniques have played an increasingly crucial role in analytical chemistry, due to the numerous advantages they offer. Several of these techniques (e.g., Near-InfraRed—NIR—or Fourier Transform InfraRed—FT-IR—spectroscopy) are considered particularly valuable because, by means of suitable equipment, they enable a fast and non-destructive sample characterization. This aspect, together with the possibility of easily developing devices for on- and in-line applications, has recently favored the diffusion of such approaches especially in the context of foodstuff quality control. Nevertheless, the complex nature of the signal yielded by spectroscopy instrumentation (regardless of the spectral range investigated) inevitably calls for the use of multivariate chemometric strategies for its accurate assessment and interpretation. This review aims at providing a comprehensive overview of some of the chemometric tools most commonly exploited for spectroscopy-based foodstuff analysis and authentication. More in detail, three different scenarios will be surveyed here: data exploration, calibration and classification. The main methodologies suited to addressing each one of these different tasks will be outlined and examples illustrating their use will be provided alongside their description.

Biological influence of cry1Ab gene insertion on the endophytic bacteria community in transgenic rice

The commercial release of genetically modified (GMO) rice for insect control in China is a subject of debate. Although a series of studies have focused on the safety evaluation of the agroecosystem, the endophytes of transgenic rice are rarely considered. Here, the influence of endophyte populations and communities was investigated and compared for transgenic and nontransgenic rice. Population-level investigation suggested that cry1Ab gene insertion influenced to a varying degree the rice endophytes at the seedling stage, but a significant difference was only observed in leaves of Bt22 (Zhejiang22 transgenic rice) between the GMO and wild-type rice. Community-level analysis using the 16S rRNA gene showed that strains of the phyla Proteobacteria and Firmicutes were the predominant groups occurring in the three transgenic rice plants and their corresponding parents. By contrast, the endophytic communities of Minghui63 and Xiushui11 showed a weaker response to cry1Ab gene insertion than did Zhejiang22, and the community results were consistent with the population-level investigation. The populations and communities of rice endophytes were affected by the cry1Ab gene to a different extent in different rice varieties and plant tissues. The results of this study broaden our understanding of unexpected transgenic influences on nontarget organisms.

Analyzing Raman spectroscopic data

This chapter is a short introduction into the data analysis pipeline, which is typically utilized to analyze Raman spectra. We empathized in the chapter that this data analysis pipeline must be tailored to the specific application of interest. Nevertheless, the tailored data analysis pipeline consists always of the same general procedures applied sequentially. The utilized procedures correct for artefacts, standardize the measured spectral data and translate the spectroscopic signals into higher level information. These computational procedures can be arranged into separate groups namely data pre-treatment, pre-processing and modeling. Thereby the pre-treatment aims to correct for non-sample-dependent artefacts, like cosmic spikes and contributions of the measurement device. The block of procedures, which needs to be applied next, is called pre-processing. This group consists of smoothing, baseline correction, normalization and dimension reduction. Thereafter, the analysis model is constructed and the performance of the models is evaluated. Every data analysis pipeline should be composed of procedures of these three groups and we describe every group in this chapter. After the description of data pre-treatment, pre-processing and modeling, we summarized trends in the analysis of Raman spectra namely model transfer approaches and data fusion. At the end of the chapter we tried to condense the whole chapter into guidelines for the analysis of Raman spectra.

Rapid identification and quantification of Panax notoginseng with its adulterants by near infrared spectroscopy combined with chemometrics

Traditional methods for identification of Panax notoginseng (PN) such as high performance liquid chromatography (HPLC) and gas chromatography (GC) are time-consuming, laborious and difficult to realize rapid and online analysis. In this research, the feasibility of identification and quantification of PN with rhizoma curcumae (RC), Curcuma longa (CL) and rhizoma alpiniae offcinarum (RAO) are investigated by using near infrared (NIR) spectroscopy combined with chemometrics. Five chemical pattern recognition methods including hierarchical cluster analysis (HCA), partial least squares-discriminant analysis (PLS-DA), artificial neural networks (ANN), support vector machine (SVM) and extreme learning machine (ELM) are used to build identification model of the dataset with 109 samples of PN and its three adulterants. Then seven datasets of binary, ternary and quaternary adulterations of PN are designed, respectively. Five multivariate calibration methods, i.e., principal component regression (PCR), support vector regression (SVR), partial least squares regression (PLSR), ANN and ELM are used to build quantitative model and compared for each dataset, separately. Finally, in order to further improve the prediction accuracy, SG smoothing, 1st derivative, 2nd derivative, continuous wavelet transform (CWT), standard normal variate (SNV), multiple scatter correction (MSC) and their combinations are investigated. Results show that PLS-DA and SVM can achieve 100% classification accuracy for identification of 109 PN with its three adulterants. PLSR is an optimal calibration method by comprehensive consideration of prediction accuracy, over-fitting and efficiency for the quantitative analysis of seven adulterated datasets. Furthermore, the predictive ability of the PLSR model for PN contents can be improved obvious by pretreating the spectra by the optimal preprocessing method, with correlation coefficients of which all higher than 0.99.

Simultaneous classification of multiple classes in NMR metabolomics and vibrational spectroscopy using interval-based classification methods: iECVA vs iPLS-DA

Interval based chemometric algorithms have proven to be very powerful for spectral alignments, spectral regressions and spectral classifications. The interval-based methods may not only improve the performance, but also reduce model complexity and enhance the spectral interpretation. Extended Canonical Variate Analysis (ECVA) is a powerful method for multiple group classifications of multivariate data and can easily be extended to an interval approach, iECVA. This study outlines the iECVA method and compares its performance to interval Partial Least Squares Discriminant Analysis (iPLS-DA) on three spectroscopic datasets from Nuclear Magnetic Resonance (NMR), Near Infrared (NIR) and Infrared (IR) spectroscopy, respectively. The results invariantly show that the interval-based classification methods greatly enhance the interpretability of the models by identifying important spectral regions, which facilitate interpretation and biomarker discovery. Although the results for the two methods are similar regarding the number of misclassifications and identified important regions, the model complexity of the PLS-DA proved to consistently lower than the ECVA. The Matlab source codes for both iECVA and iPLS-DA are made freely available at www.

MODELS

life.ku.dk.

Classification and quantitation of milk powder by near-infrared spectroscopy and mutual information-based variable selection and partial least squares

Milk is among the most popular nutrient source worldwide, which is of great interest due to its beneficial medicinal properties. The feasibility of the classification of milk powder samples with respect to their brands and the determination of protein concentration is investigated by NIR spectroscopy along with chemometrics. Two datasets were prepared for experiment. One contains 179 samples of four brands for classification and the other contains 30 samples for quantitative analysis. Principal component analysis (PCA) was used for exploratory analysis. Based on an effective model-independent variable selection method, i.e., minimal-redundancy maximal-relevance (MRMR), only 18 variables were selected to construct a partial least-square discriminant analysis (PLS-DA) model. On the test set, the PLS-DA model based on the selected variable set was compared with the full-spectrum PLS-DA model, both of which achieved 100% accuracy. In quantitative analysis, the partial least-square regression (PLSR) model constructed by the selected subset of 260 variables outperforms significantly the full-spectrum model. It seems that the combination of NIR spectroscopy, MRMR and PLS-DA or PLSR is a powerful tool for classifying different brands of milk and determining the protein content.

Predictive modelling of the water contact angle of surfaces using attenuated total reflection – Fourier transform infrared (ATR-FTIR) chemical imaging and partial least squares regression (PLSR)

A simple method to model and predict the surface wettability using a combination of ATR-FTIR imaging and PLSR modelling.

Measuring Biological Age via Metabonomics: The Metabolic Age Score

Chronological age is one of the most important risk factors for adverse clinical outcome. Still, two individuals at the same chronological age could have different biological aging states, leading to different individual risk profiles. Capturing this individual variance could constitute an even more powerful predictor enhancing prediction in age-related morbidity. Applying a nonlinear regression technique, we constructed a metabonomic measurement for biological age, the metabolic age score, based on urine data measured via (1)H NMR spectroscopy. We validated the score in two large independent population-based samples by revealing its significant associations with chronological age and age-related clinical phenotypes as well as its independent predictive value for survival over approximately 13 years of follow-up. Furthermore, the metabolic age score was prognostic for weight loss in a sample of individuals who underwent bariatric surgery. We conclude that the metabolic age score is an informative measurement of biological age with possible applications in personalized medicine.