The use of chemometrics to study multifunctional indole alkaloids from Psychotria nemorosa (Palicourea comb. nov.) was developed by Júnior et al.³. The techniques PCA, partial least squares (PLS) and orthogonal PLS1 (O-PLS1)were helpful for modelling the activities using ultra-high-performance liquid chromatography-diode array detector (UPLC-DAD) data as fingerprint³. Ultraviolet spectral data from monoclonal antibodies compounded into hospital pharmacy were evaluated to build models using PLS-discriminant analysis (PLS-DA). The challenge was to identify the monoclonal antibodies after compounding and just before administration to the patient for quality control at the hospital⁴. The linear sweep voltammograms generated from integration of three commercial screen-printed electrodes into a voltammetry electronic tongue was combined with PLS to predict concentrations of cysteine, glutathione and homocysteine. The authors applied genetic algorithm and interval PLS (i-PLS) for selection of variables⁵. It was possible to monitor the extraction of indole alkaloids from the toad skin⁶ and the quantification of geniposide in Gardenia jasminoides fruit (Chinese medicinal herb)⁷ was performed using near infrared (NIR) spectroscopy and PLS regression (PLSR). The extraction of indole alkaloids from the toad skin showed a coefficient of determination (R²) value of 0.99 and root mean square error of cross validation (RMSECV) of 8.26 mg mL^{-1 6} and the correlation values to quantify the concentration of geniposide in Gardenia were between 0.92 – 0.99 and 0.32 - 1.66 mg mL^-1 for RMSECV⁷. In order to evaluate the anti-inflammatory properties of extracts from Honeysuckle, Fourier Transform Infrared (FTIR) spectroscopy in combination with PLSR⁸ were used.

Junior et al.⁹ used NIR and performed a comparison between two multivariate calibration techniques, PLSR and multiple linear regression (MLR) to monitor the quality control of tablets, and the PLSR model present better results than MLR. Raman spectral imaging technique was also used to predict active ingredient pharmaceutical concentration in microtablets¹⁰, in oral suspensions¹¹, intravaginal rings¹², and in quantitative assessment of pharmaceutical tablets¹³. The PLSR and least squares-support vector machine (LS-SVM) models exhibited excellent prediction abilities for active pharmaceutical ingredient in microtablets with a correlation higher than 0.98 and root mean square error of prediction (RMSEP) of 2.35% (w/w) for PLSR and 2.09% (w/w) for LS-SVM¹⁰. Mazurek and Szostak¹¹, using PLSR model, obtained RMSE of calibration (RMSEC) from 1.59 to 2.92% for oral suspensions¹¹ and Lyndgaard et al.¹² achieved RMSECV of 1.82% (w/w) and correlation of 0.99 for intravaginal rings using PLSR model. Then, Sparén et al.¹³ in quantitative assessment of tablets obtained RMSE of prediction (RMSEP) from 0.54 to 3.04% (w/w). The ultraviolet visible (UV-Vis) spectroscopy and chemometric models also allowed determination the presence of impurities in tablets¹⁴. Support vector regression (SVR) model was more accurate when compared to PLSR model with values for RMSEP between 0.18 – 0.27 mg mL^-1 while the values for PLSR were RMSEP from 0.27 to 0.30 mg mL^{-1 14}. The synergy PLS (siPLS) model combined with FTIR showed the best results compared to the full spectrum PLS model for the quantification of active ingredient pharmaceutical in tablets which values highlight by authors was RMSECV of 15.28 mg g^-1 and R² of 0.99 for sulphamethoxazole and RMSECV of 7.52 mg g^-1 and correlation of 0.98 for trimethoprim¹⁵. Cantor et al.¹⁶ evaluated the ability of different analytical techniques such X ray diffraction, FTIR, Raman and NIR to detect melamine levels in gelatin (pharmaceutical component). The authors highlighted NIR technique and PLSR model that achieved the best results with RMSEP of 2.5% (w/w) ¹⁶, besides to quantify the presence of melamine at levels as low as 1.0% (w/w).

The NIR spectroscopy and PLSR were used to determine protein nitrogen in yellow fever vaccine¹⁷. Haghighi, Hemmateenejad and Shamsipur¹⁸ using fluorescence spectroscopy and multivariate curve resolution alternative least square (MCR-ALS) were able to predict enantiomeric excess of some amino acids in frozen plasma which RMSEP values lower than 0.07% and R higher than 0.97. PLS was applied in NIR spectra to analyze the indicator of β-thalassemia in human hemolysate samples¹⁹. FTIR and multivariate calibration method were used to identify three group of clinical bacterial pathogens in human serum²⁰. Both models, PCA and PLS-DA, revealing a feasible tool for discrimination of the samples with percent accuracy of 100%²⁰.

NIR and PLSR were used to measure both physicochemical²¹ and the antioxidant properties of honey²², as well to detect and to quantify adulterant in honey samples of stingless bees²³. The PLSR model presented a RMSECV between 0.008 – 1.47 and correlation values from 0.90 to 0.98 for physicochemical properties²¹, while the antioxidant properties of honeys from different botanical origin showed high correlation varying from of 0.95-0.96 and RMSECV range from 0.92 to 13.60 ²². Se et al.²³ obtained values for prediction of adulteration in honey samples of stingless bees with relative errors up 0.58 to 1.49%. The three-dimensional fluorescence spectroscopy technique combined with multivariate calibration was also used for detecting the presence of adulterants in honey²⁴.

A near infrared hyperspectral imaging system was used for predicting viability and vigor in muskmelon seeds using PLS-DA. Variable selection methods, such as variable important projection (VIP), selectivity ratio (SR) and significance multivariate correlation (sMC) were also applied to select the most effective wavelengths²⁵.

In order to determine oxidation parameters in frying canola oil, Talpur et al.²⁶ employed FTIR technique and PLSR model, while Uncu and Ozen²⁷ studied various chemical parameters in olive oils and Wu et al. evaluated rapeseed oil adulteration²⁸. NIR and PLSR models were also used to study oxidation of edible oils²⁹, to detect adulteration in extra virgin olive oils³⁰, and to monitor vegetable oils quality³¹. The fluorescence spectroscopy and different multivariate calibration methods, PLSR, PCR and MLR were applied for detection of adulteration of sesame oil. The best results highlighted by the authors were using the PLSR model which figures of merit were RMSECV of 0.65 for 80 nm wavelength intervals³².

The applicability of FTIR was used as tool to determine of total sugar content in soy-based drinks³³ and for measuring the concentration of nutrients in grapevine petioles³⁴. For the soy-based drinks, two models were applied for different spectral band: (1) iPLS and (2) siPLS, and both presented high coefficient of correlation (R) (0.98).

Milk samples were evaluated using NIR and PLS-DA to detect contamination by Salmonella³⁵. Mid-infrared (MIR) spectroscopy and PLSR were applied for determining the residues of tetracycline in cow’s milk³⁶ and for detecting and quantifying the adulteration in milk powder³⁷. Augusto et al. used laser-induced breakdown spectroscopy (LIBS) data and PLSR model for the direct determination of mineral elements, such as Ca, K and Mg in powdered milk and dietary supplements³⁸.

Data acquired using C-13 nuclear magnetic resonance (NMR) spectroscopy were evaluated using SVR with variable selection by genetic algorithm (GA) to determine the contents of saturated, aromatic, and polar compounds in crude oil. The use of small amounts of samples was highlighted by the authors³⁹. Sulfur content in petroleum derivatives was determined using MIR spectroscopy and PLS in association with variable selection methods, among them iPLS, siPLS, uninformative variable elimination (UVE), and GA⁴⁰. The UV spectroscopy integrated different chemometric strategies for quantifying the hydrocarbon contents in fuel oil samples⁴¹.

In order to quantify the presence of adulterants in gasoline, Raman spectroscopy was used, and calibration models were constructed using PLSR, iPLS and PLS-GA which correlations were between 0.80 and 0.99⁴². Gas chromatography data were used to build a PLSR model for quality assessment of gasoline, and low errors from 0.005 to 0.010% were obtained⁴³. Dadson, Pandam and Asiedu evaluated adulterants in gasoline, such as kerosene, diesel, naphta using a multivariate model with FTIR data and PLSR⁴⁴. Mabood et al.⁴⁵ also applied PLSR model to detect and estimate gasoline adulteration by NIR spectroscopy.

Neves et al.⁴⁶ used mass spectrometry coupled with PCA, GA with support vector machine (SVM), linear discriminant analysis (LDA) and quadratic discriminant analysis (QDA) as an untargeted lipidomic approach to classify 76 blood plasma samples into negative for intraepithelial lesion or malignancy and squamous intraepithelial lesion. The PCA-SVM models provided the best classification results, achieving values of 80% and 83% for the sensitivity and specificity, respectively. Santos et al.⁴⁷ proposed a method to determine 11 polycyclic aromatic hydrocarbons (PAHs) in urine samples based on the coupling of a programmed temperature vaporizer (PTV) with a quadrupole mass spectrometer (qMS) instrument, via a deactivated fused silica tubing. The authors used PLS-DA, LDA, soft independent modeling of class analogy (SIMCA) and SVM to classify the samples according to the presence or absence of the PAHs. Gilany et al.⁴⁸ investigated untargeted metabolomic profiling of the seminal plasma in non-obstructive azoospermia men using gas chromatography-mass spectrometry (GC-MS) and QDA chemometric technique to implement on total ion chromatograms for identification of discriminatory retention times. The receiver operating characteristic (ROC) curves for these classification models presented 88% of accuracy for the discrimination of 36 metabolites that may be considered discriminatory biomarkers for different groups in non-obstructive azoospermia. Boll et al.⁴⁹ developed a method using attenuated total reflection (ATR) FTIR spectroscopy combined with PLS-DA model that allowed the discrimination between dyed and non-dyed hair of 380 hair samples. Monakhova, Diehl and Fareed⁵⁰ used PCA, factor discriminant analysis (FDA), PLS-DA and LDA combined with high resolution (600 MHz) NMR spectroscopy data to distinguish 102 authentic samples of heparin and low-molecular weight heparins produced from porcine, bovine and ovine mucosal tissues as well as their blends.

Chen et al.⁵¹detected 101 volatile compounds in 70 Chinese vinegars through GC-MS and used chemometric techniques such as clusters analysis, LDA, k nearest neighbor (kNN), GA-ANN to classify these compounds. Giannetti et al.⁵²evaluated flavor composition of apple cultivars grown in the Northeast Italy through different cultivation forms. The authors used head space-solid phase micro extraction/gas chromatography mass spectrometry (HS-SPME/GC-MS) for the analysis of volatile fraction and PLS-DA model to classify 42 apples varieties. The HS-SPME/GC-MS technique identified 118 volatile compounds and the PLS-DA model classified apples based on their different geographical origin or growing conditions, providing accuracy of 88-100%. Cuevas et al.⁵³ applied data fusion (mid-level) using high-performance liquid chromatography high resolution mass spectrometric (HPLC-HR-MS) and HS-SPME/GC-MS to obtain data sets and PLS-DA model for the discrimination between 19 samples of organic and conventional orange juices. The HPLC-HR-MS and HS-SPME/GC-MS techniques identified some flavonoids, fatty acids, aldehydes and esters as potential markers involved in the discrimination of organic juices. Wen et al.⁵⁴ characterized the phenolic and proline profiles of four types of unifloral honey collected from beekeepers in China. LDA results showed overall accuracy of 99% and 94% in correct classification of differentiate the floral origin of four Chinese honeys. Portarena, Baldacchini and Brugnoli⁵⁵ proposed a method to discriminate 38 extra-virgin olive oils from seven regions along the Italian coasts from determination of the isotopic composition and the carotenoid contents. The authors used isotope ratio mass spectrometry (IRMS) and resonant Raman spectroscopy (RRS) to determine isotopic composition and the carotenoid contents, respectively. The LDA model presented 50 – 100% of correct classification for the seven regions, achieving 82% of accuracy for validation set.

Márquez et al.⁵⁶used NIR and FT-Raman data to perform the data fusion in mid and high-level using SIMCA model for classifying the unadulterated and adulterated hazelnut samples. Rodrigues Júnior et al.⁵⁷ applied PLS-DA for discrimination of lactose-free samples and classification of adulterated and unadulterated milk powder samples by FT-Raman spectroscopy. Wakholi et al.⁵⁸used short-wave IR (SWIR) hyperspectral imaging coupled with LDA, PLS-DA and SVM models for classification of corn seeds viable (control group) and corn seeds non-viable (microwave treated). The authors used a total of 200 samples of each type of corn: white, yellow and purple. Sandasi et al.⁵⁹applied a fast and non-destructive method for the quality control of herbal tea blends. The authors used hyperspectral images obtained from SWIR combined with PLS-DA to classify herbal tea blends of the species Sceletium tortuosum and Cyclopia genistoides.

Kimuli et al.⁶⁰ also used SWIR hyperspectral imaging system (1000 – 2500 nm) to detect aflatoxin B₁ combined with PCA, PLS-DA and FDA techniques to explore and classify maize kernels of four varieties from different States of the USA. Cortés et al.⁶¹ discriminated two nectarine varieties in different spectral ranges of NIR and Vis-NIR using PLS-DA and LDA models. Shrestha et al.⁶² used NIR hyperspectral imaging data from 975 to 2500 nm and PLS-DA to investigate seed quality parameters such as year of production and variety in tomato seed lots. Santos, Pereira-Filho and Colnago⁶³ applied ¹H time domain nuclear magnetic resonance (TD-NMR) combined with multivariate analysis for identifying and quantifying the adulteration in 78 samples of milk using whey, urea, hydrogen peroxide and synthetic material (urine or milk).

Chen et al.⁶⁴ used NIR combined with hierarchical cluster analysis (HCA), PLS-DA and SVM to authenticate stamps of 12 seals on a Chinese traditional painting. The results obtained were 93% and 100% of accuracy for the PLS-DA and SVM model in validation set, respectively. Oliveira et al.⁶⁵ discriminated banknotes authenticity of Brazilian Real (R$ 20, R$ 50 and R$ 100) using a portable NIR spectrometer combined with SIMCA and successive projections algorithm-LDA (SPA-LDA) models. Martins et al.⁶⁶ proposed a method for direct analysis of seven whisky brands. The authors used UV-Vis spectroscopy combined with PLS-DA to classify 164 genuine and 73 false samples. The PLS-DA corrected classified 99% and 93% for genuine and false samples, respectively. Costa et al.⁶⁷ identified six polymers using different ratios of emission lines and molecular bands (C/H, C/C₂, C/N, C/O, H/C₂, H/N, H/O and O/N) obtained with LIBS. SIMCA and kNN models were computed using 477 samples divided between calibration and validation sets. Both models achieved accuracy of 98% (kNN) and 92% (SIMCA).

Silva et al.⁶⁸ developed a method to classify gasoline according to its origin (Brazil, Venezuela and Peru), using IR spectroscopy and multivariate classification. A set of 126 gasoline samples: 56 Brazilian, 66 Venezuelan, and 4 Peruvian, was analyzed. The spectra were standardized using the direct standardization method achieving 100% of correct predictions. In another study, Silva et al.⁶⁹ proposed the use of IR spectroscopy and LDA and PLS-DA for classification of gasoline with or without dispersant and detergent additives in 125 samples. The authors selected the variables for the multivariate models using stepwise (SW), successive projections algorithm (SPA) and GA algorithms. Sinkov, Sandercock and Harynuk⁷⁰ developed PLS-DA and SIMCA models for classifying the levels of gasoline in casework arson samples based on GC-MS data.

fabiola.verbi@unesp.br