Chemometrics and Intelligent Laboratory Systems最新文献

The fast and accurate prediction of Hansen solubility benefits many diverse fields such as pharmaceuticals, the food industry, and cosmetics. To estimate the individual HSP values (polar, dispersive, and hydrogen bonding components), we investigated the performance of using Mordred descriptors in multiple linear regressions and XGBoost modeling. For HSP predictions, we also tested a graph-based molecular representation with graph neural network (GNN) modeling. To select the optimal models for final training and predictions, we used nested cross-validation and hyper-parameter optimization. The models with the best predictive performance were selected through internal (R²_train, RMSE, MEPcv) and external (RMSEP, CCC, MEP, R²_test, ar²m, Δr²m) validation metrics using ∼1200 compounds from free-available database https://www.stevenabbott.co.uk. To confirm the practical reliability, we examined the agreement of experimentally obtained HSP data from the literature for 93 compounds and the data predicted by the created models. The results of GNN modeling showed the best predictive characteristics, which include a coefficient of determination between experimentally obtained and predicted HSP values greater than 0.76 for polar and hydrogen bond forces and greater than 0.66 for dispersive forces. Interpreting the fundamental basis of Hansen solubility using the created MLR equations and XGBoost models, HSP values were found to be influenced by van der Waals volume characteristics, 2D matrix molecular representation, and polarity. We elaborated on the practical benefits of using the selected GNN method through Hansen's solubility sphere as an example. This is the first study to demonstrate the advantages of GNN in predicting individual HSP components, as well as the first study to describe in detail their molecular basis using MLR and XGBoost modeling.

LipidQuant 2.1 is a software written in Matlab, which is designed for the high-throughput processing of large lipidomic data sets measured by lipid class separation coupled with quadrupole time-of-flight (QTOF) high-resolution mass spectrometry (MS). The software enables the identification of lipid species based on defined mass accuracy. The main focus is on the right lipidomic quantitation using at least one internal standard per lipid class and the implementation of an automated procedure for Type I and Type II isotopic corrections necessary for the determination of accurate molar concentrations, which is not available for the majority of existing software solutions. LipidQuant 2.1 offers three options for peak assignment, visualization of the isotopic pattern, and automated calculation of m/z for various adduct ions. The initial lipidomic database covers 31 lipid classes with more than 2900 lipid species that occur primarily in the human lipidome, but users have the full flexibility to modify and extend the database according to their needs. All algorithms and the detailed user manual are provided. The reliability of LipidQuant 2.1 is demonstrated on a set of more than 250 biological samples measured by ultrahigh-performance supercritical liquid chromatography (UHPSFC) coupled with QTOF-MS.

The paper deals with the inversion of intervals when a PLS (Partial Least Squares) model is used. However, instead of discretizing the interval, it is proved that the region resulting from the inversion of a PLS model is a convex set bounded by two parallel hyperplanes, each corresponding to the direct inversion of each endpoint of the given interval.

When the domain of the input variables is a convex set, any feasible solution with predictions within the interval set in the response can be obtained as a convex combination of a point on each of the two hyperplanes. In this way, the new solutions preserve the internal structure of the input variables.

This methodology can be of interest in several domains where the response under study is defined in terms of an interval of admissible values, such as specifications for a product in an industrial process, or tolerance intervals for computing compliant class-models.

The inversion of the corresponding fitted model defines a region in the input space (predictor variables) whose predictions fall within the specified interval. Then, estimating and exploring this region will increase the information about the problem under study.

Here is presented a procedure that extends standard PLS Regression to several data matrices in a path. The basic idea is to convert the path of data matrices into interconnected regressions. Forecasts by PLS are extended to multi-step forecasts for each data matrix in the path. We study how far we can make forecasts, i.e., how far we can ‘see’ in the path. It is shown how data paths are divided into parts, where multi-step forecasting can be carried out within each part. The principles of PLS are used to suggest criteria for estimation in the regressions. These methods can be used to supervise a complex path of industrial chemical/biological processes. It is shown how expanding and contracting paths, which is common for industrial processes, can be handled. These methods can be used to carry out analysis of general path models. It is shown briefly by an example how a Structural Equations Model, SEM, can be converted into a collection of sequential paths that can be analyzed by present methods. The results suggest that conclusions made at SEM analysis may not always be reliable. The theory is applied to process data. It is shown how we work with the analysis of each regression in a similar way as in PLS.

In this study, 110 tea samples from South American countries (Argentina, Brazil, and Paraguay) and Asian countries (India and China) were analyzed using near-infrared spectroscopy (NIRS) together with a two-step chemometric authentication strategy (class modeling techniques and discriminant analysis) to authenticate commercial teas from Argentina. In the first step, one-class models were built and validated to authenticate South American teas using preprocessed NIRS data. For this purpose, data-driven soft independent modeling of class analogy (DD-SIMCA) and one-class partial least squares (OC-PLS) were used. The DD-SIMCA model gave the best results, with a sensitivity of 93.10%, specificity of 100%, and efficiency of 95.00%. In the second step, a support vector machine (SVM) was used to build and validate a multiclass model to discriminate between tea samples from Argentina and neighboring countries of South America. The best model was the combination of nine variables selected by the fast correlation-based filter (FCBF) method, with an accuracy of 98.30%. Therefore, we conclude that the combination of NIRS and two-step chemometric tools can be used to authenticate the geographical origin of samples with high inter-class similarity.

Unbalanced sample groups tend to yield models with a higher prevalence of predominant classes. A sample group with balanced classes contributes to the development of more robust models with improved predictive capability to classify classes equally. In the literature, two methodologies for sample balancing can be found: elimination (undersampling) and synthetic sample generation (oversampling). Undersampling methodologies result in the loss of real samples, while oversampling methods may introduce issues related to adding non-real signals to the original spectra. To overcome these challenges, this paper aimed to utilize Principal Component Analysis (PCA) for the generation of virtual samples (synthetic samples and artificial outliers) to balance data in multivariate classification models. The proposed methodology was applied to data from mid-infrared spectroscopy (MIR) and high-resolution mass spectrometry (HRMS) with Partial Least Squares Discriminant Analysis (PLS-DA) and Support Vector Machine (SVM) models. The constructed models demonstrate that the addition of virtual samples enhances performance parameters (e.g., false negative rate, false positive rate, accuracy, sensitivity, specificity, among others) compared to unbalanced models, while also mitigating overfitting (a problem found in unbalanced models). Performance parameters exhibited a more significant improvement percentage using the non-linear model (SVM) compared to the linear model (PLS-DA). Furthermore, the created virtual spectra do not introduce new signals, i.e., original, and virtual spectra exhibit a similar spectral profile, differing only in the intensity levels. Finally, all models demonstrated good predictive capability according to permutation testing for the binary model developed in this work, limiting the rate of class permutation retention (between 40 % and 60 % of the y-vector remained in the original class). All created models exhibited accuracy values higher than the accuracy distribution of models with permuted classes for the test group.

Dealing with missing data poses a challenge in Principal Component Analysis (PCA) since the most common algorithms are not designed to handle them. Several approaches have been proposed to solve the missing value problem in PCA, such as Imputation based on SVD (I-SVD), where missing entries are filled by imputation and updated in every iteration until convergence of the PCA model, and the adaptation of the Nonlinear Iterative Partial Least Squares (NIPALS) algorithm, able to work skipping the missing entries during the least-squares estimation of scores and loadings. However, some limitations have been reported for both approaches. On the one hand, convergence of the I-SVD algorithm can be very slow for data sets with a high percentage of missing data. On the other hand, the orthogonality properties among scores and loadings might be lost when using NIPALS.

To solve these issues and perform PCA of data sets with missing values without the need of imputation steps, a novel algorithm called Orthogonalized-Alternating Least Squares (O-ALS) is proposed. The O-ALS algorithm is an alternating least-squares algorithm that estimates the scores and loadings subject to the Gram-Schmidt orthogonalization constraint. The way to estimate scores and loadings is adapted to work only with the available information.

In this study, the performance of O-ALS is tested and compared with NIPALS and I-SVD in simulated data sets and in a real case study. The results show that O-ALS is an accurate and fast algorithm to analyze data with any percentage and distribution pattern of missing entries, being able to provide correct scores and loadings in cases where I-SVD and NIPALS do not perform satisfactorily.

Functional MRI is, currently, the most sensitive technique in breast cancer for detecting early tumors, and perfusion (DCE-MRI) has become the most important sequence to depict and characterize angiogenesis and neovascularization. In this work, we propose the use of new biomarkers that are related to clear physiological phenomena, obtained from MCR-ALS as an alternative to curve-based pseudo-biomarkers and pharmacokinetics models. In order to provide a discrimination and prediction model between healthy tissue and cancer, we propose using PLS-DA with double cross-validation (2CV) and variable selection, repeated several times and obtaining excellent average results for the performance indexes (f-score: 0.9149, MCC: 0.8538, AUROC: 0.8794). After selecting the optimal prediction model, a unique probabilistic map called “virtual biopsy” that shows in different colors the probability that each pixel of the image has a tumor behavior is obtained, helping the specialist with the identification and characterization of breast tumors with only one easy-to-interpret biomarker map.

Time-series prediction is of great practical value in industrial scenarios such as rotary kilns, especially for long sequence time-series prediction. Accurate long sequence NOx emission predictions help us monitor rotary kiln operations in advance to plan and control NOx emissions according to emission policies and production requirements. However, in actual industrial scenarios, the NOx emission pattern is dominated by long-term trends rather than simply repetitive patterns. Existing NOx prediction models are not effective in capturing long-term dependencies. Therefore, this paper proposes a novel model based on Transformer to solve this problem. First, we propose a novel series decomposition architecture based on LSTM and self-attention, which is embedded inside the Transformer. The architecture allows self-attention at the sub-series level and provides short-term trend and position information. In addition, the model designs a one-step inference structure to improve the error accumulation phenomenon under traditional inference methods for long sequence prediction and reduce the inference time. We conducted extensive experiments on two real-world datasets with different sampling intervals, which validated the model’s effectiveness. It achieves a relative improvement of 53.2% and 43.4% in prediction accuracy compared to popular NOx emission prediction methods.