Journal of Chemometrics最新文献

Angiogenesis, a crucial process in tumor growth, is widely recognized as a key factor in cancer progression. The vascular endothelial growth factor (VEGF) signaling pathway is important for its pivotal role in promoting angiogenesis. The primary objective of this study was to identify a powerful classifier for distinguishing compounds as active or inactive inhibitors of VEGF receptors. To build the machine learning model, compounds were sourced from the BindingDB database. A variety of common feature selection techniques, including both filter-based and wrapper-based methods, were applied to reduce dimensionality, subsequently, overfitting problem. Robust and accurate tree-based classifiers were employed in the classification procedure. Application of the extra-tree classifier using the MultiSURF* feature selection method provided a model with superior accuracy (83.7%) compared with other feature selection techniques. High-throughput molecular docking followed by an accurate docking and comprehensive analysis of the results was performed to provide the best possible inhibitors of these receptors. Comprehensive analysis of the docking results revealed successful prediction of molecules with VEGFR1 and VEGFR2 inhibitory activity. These results emphasized that the performance of the extra-tree model, coupled with MultiSURF* feature selection, surpassed other methods in identifying chemical compounds targeting specific VEGF receptors.

Real-time selective protein quantification is an integral component of operating continuous chromatography processes. Partial least squares models fit with spectroscopic UV-Vis absorbance data have demonstrated the ability to selectively quantify proteins. With standard continuous chromatography equipment that is only capable of measuring absorbance at a few user-defined wavelengths, the problem of selecting appropriate wavelengths that maximize the measurement capability of the instrument remains unaddressed. Therefore, we propose a method for selecting wavelengths for continuous chromatography equipment. We illustrate our method using sets of protein mixtures composed of bovine serum albumin and lysozyme. The first step is to refine the raw wavelength set with a statistical t-test and an absorbance magnitude test. Then, the wavelengths within the refined spectroscopic range are ranked. Three existing techniques are evaluated – sequential forward search, variable importance to projection scores, and the least absolute shrinkage and selection operator. The best technique (in this case, sequential forward search) determines a subset of three wavelengths for further evaluation on the BioSMB PD. We use an exhaustive approach to determine the final wavelength set. We show that soft sensor models trained from the method's wavelength selections can quantify the two proteins more accurately than from the wavelength set of 230, 260 and 280 nm, by a factor of four. The method is shown to determine appropriate wavelengths for different path lengths and protein concentration ranges. Overall, we provide a tool that alleviates the analytical bottleneck for practitioners seeking to develop advanced monitoring and control methods on standard equipment.

The t-distributed stochastic neighbour embedding algorithm or t-SNE is a non-linear dimension reduction method used to visualise multivariate data. It enables a high-dimensional dataset, such as a set of infrared spectra, to be represented on a single, typically two-dimensional graph, revealing its global and local structure. t-SNE is very popular in the machine learning community and has been applied in many fields, generally with the aim of visualising large datasets. In vibrational spectroscopy, t-SNE is gaining notoriety but principal component analysis (PCA) remains by far the reference method for exploratory analysis and dimension reduction. However, t-SNE may represent a real aid in the analysis of vibrational spectroscopic datasets. It provides an at-a-glance global view of the dataset allowing to distinguish the main factors influencing the spectral signal and the hierarchy between these factors, and gives an indication on the possibility of performing predictive modelling. It can also provide great support in the choice of the pre-processing, by comparing rapidly different general pre-processing approaches according to their effect on the variable of interest. Here we propose to illustrate these advantages using different datasets. We also propose an approach based on a synergy between the t-SNE and PCA methods, allowing respective advantages of each to be exploited.

This study illustrates at-line application of hyperspectral imaging in the visible range for quality control of large-scale offset printing. In particular, the measurement stability of a competitive device is assessed and compared to traditional handheld and desktop spectrophotometers. The performance of the commercially available instruments was assessed based on collected spectra and their corresponding L*, a*, and b* values. The printing process was described by hyperspectral images (in visible range) of selected regions from template color fields acquired at 17 sampling occasions. Spectra constituting hyperspectral images were visualized and evaluated in the space of significant principal components obtained from the principal component analysis. Furthermore, confidence ellipses were constructed for each set of spectra characterizing a specific moment of the printing process. Comparing their mutual locations, shapes, orientations, and sizes enabled effective visualization of process variability and was more comprehensive regarding the classic approach based on information provided by desktop and handheld spectrometers.

In this project, we used formalin-fixed paraffin-embedded (FFPE) tissue samples to measure thousands of spectra per tissue core with Fourier transform mid-infrared spectroscopy using an FT-IR imaging system. These cores varied between normal colon (NC) and colorectal primer carcinoma (CRC) tissues. We created a database to manage all the multivariate data obtained from the measurements. Then, we applied classifier algorithms to identify the tissue based on its yielded spectra. For classification, we used the random forest, a support vector machine, XGBoost, and linear discriminant analysis methods, as well as three deep neural networks. We compared two data manipulation techniques using these models and then applied filtering. In the end, we compared model performances via the sum of ranking differences (SRD).

Modeling near-infrared (NIR) spectral data to predict fresh fruit properties is a challenging task. The difficulty lies in creating generalized models that can work on fruits of different cultivars, seasons, and even multiple commodities of fruit. Due to intrinsic differences in spectral properties, NIR models often fail in testing, resulting in high bias and prediction errors. One current solution for achieving generalized models is to use large calibration sets measured over multiple cultivars and harvest seasons. However, current practice primarily focuses on calibration sets for single fruit commodities, disregarding the rich information available from other fruit commodities. This study aims to demonstrate the potential of locally weighted partial least-squares an example of just-in-time (JIT) modeling to develop real-time models based on calibration sets consisting of multiple fruit commodities. The study also explores JIT modeling for leveraging relevant information from other fruit commodities or adapting the model based on new samples. The application demonstrated here predicts the dry matter in fresh fruit using portable NIR spectroscopy. The results show that JIT modeling is particularly effective for multiple fruit commodities in a single calibration set. The JIT models achieved a root mean squared error of prediction (RMSEP) of 0.69% fresh weight (FW), while the traditional partial least squares (PLS) modeling RMSEP was 0.93% FW. JIT modeling can be particularly beneficial when the user has multiple fruit datasets and wants to combine them into a single dataset to utilize all the relevant information available.

The rapid advancement of industrialization and urbanization has led to the global problem of air pollution. Air quality can decrease due to pollutants in the air, including types of gases and particles that are carcinogenic, causing adverse health effects. Therefore, estimating the concentration of air pollutants is of great interest as it can provide accurate information about air quality with proper planning of future activities. In this manner, this study considers Istanbul, a province with a high concentration of industry, population, and vehicle traffic. Particulate matter (PM), one of the most basic air pollutants, is stated to contain microscopic solids or liquid droplets that are small enough to be inhaled and cause serious health problems. Thus, it is recommended to apply discrete wavelet transform (DWT) and deep learning method long short-term memory (LSTM) as a hybrid model to predict the concentration of PM₁₀. Using the mentioned methods, they can predict air pollution to have been developed within the scope of this study. Furthermore, the hybrid approach with LSTM by selecting the most appropriate discrete wavelet type emphasizes the difference of this study from the existing literature. The ability of these developed methods to make successful future predictions helps institutions and organizations that can take precautions on the subject to take action at the right time; in addition, the deep learning methods used contribute to the development of sustainable smart environmental systems. In today's environment when air pollution is increasing and threatening human health, any precaution that can be taken would improve the quality of life for all living things, reduce health issues and deaths caused by air pollution, and thus raise the degree of well-being. These findings might offer a reliable scientific evidence for Istanbul City's air pollution management, which can serve as an example for other regions.

Aiming at the actual industrial process background that different modes share the same system configurations and control structure, this article proposes a novel structured discriminant Gaussian graph learning for multimode process monitoring. The proposed method considers not only the sparsity of graph model but also the measurement of data variation based on a mismatched graph and the common node support between different graphical structures. The objective function involves two sets of regularization terms: the trace terms for mismatched measurements and the $�{�\ell �}_{�\mathrm{2,1}�}$-norm imposed on the union of decomposed graph matrices. Due to the introduced mismatched trace terms, the cost of matching the data points and graph models that have inconsistent class labels can be expanded, which brings more discrimination for the graph-based mode identification. While the common structure extracted by the $�{�\ell �}_{�\mathrm{2,1}�}$-norm forces the estimated graph models to have structural similarities, thus alleviating the negative influence caused by graph discrimination. Once a relatively accurate and discriminative reference graph model is obtained, the downstream test graph learning and analysis can be conducted online by employing the moving window techniques. By comparing the matched and mismatched graph-based measurements, the process mode can be identified correctly and stably. To grasp the abnormal process changes, the $�{�\ell �}_{�\mathrm{2,1}�}$-norm for the row sparsity is again applied to the graph difference matrices, the sensitive monitoring statistics and the fault isolation results can be obtained effectively. All the optimization problems in this paper can be solved using the alternating direction multiplier (ADMM) algorithm. The effectiveness of our proposed approach is illustrated by the application to a real blast furnace iron-making production process.

Monitoring of flocculation processes such as those used in downstream processing of a fermentation broth is essential for process control. One approach is to apply microscopic imaging combined with image analysis for characterizing the state of the process. In this work, we investigate and compare the use of supervised feedforward convolutional neural network (CNN) architectures to predict the process states from the image information and compare the results with the traditional alternative of characterizing flocs based on manually engineered image features guided by human expertise. From a well-defined image data set representing six process states, the objective is to establish end-to-end classification models which are accurate but at the same time learn meaningful latent variable space representations. Specifically, we evaluate three different CNN architectures with varying degrees of regularization and compare results with logistic regression models based on inputs from two different traditional feature engineering methods. By applying global average pooling as a structural regularizer to the CNN architecture, we significantly improve the generalization performance in comparison with the classification accuracies of the traditional feature engineered models. Furthermore, we show that by imposing a projection to latent structures (PLS) like regularization framework onto the CNN, it can also learn a latent variable representation that mimics the features selected by human expertise.

Spectral data describing product samples are typically composed of a large number of noisy and irrelevant wavelengths that tends to undermine the performance of multivariate predictive techniques. This paper proposes a two-phase framework that integrates a preselection wavelength step oriented by wavelength clustering to a wrapper-based strategy. The first phase performs a pruning process in the data that removes the less informative wavelengths relying on the spectral clustering, a technique deemed suitable to the Fourier transform infrared (FTIR) spectroscopy and near-infrared (NIR) spectroscopy data at hand. The preselected wavelengths undergo a second phase of selection efforts based on the combination of different wavelength importance indices (i.e., Bhattacharyya distance, Chi-square, ReliefF, and Gini) and classification techniques (i.e., support vector machine, k-nearest neighbors, and random forest). When applied to 11 FTIR datasets from different domains, the recommended combination of importance index and classifier increased the average accuracy by 6.37% (from 0.863 to 0.918), while retaining average 3.84% of the original spectra. The framework also improved the selection process regarding computational time.