Journal of Chemometrics最新文献

This study presents a novel Monte Carlo–based methodology for estimating classification uncertainty in chemometric models by propagating spectral measurement noise. Unlike traditional approaches that treat classification as deterministic, this method simulates realistic noise structures, both independent and correlated, captured from multiple spectrum measurements to quantify sample-specific uncertainty. The technique is applicable to both linear and non-linear models, including partial least squares discriminant analysis (PLS-DA) and various support vector machine (SVM) kernels. The methodology was validated using three datasets: synthetic 2D simulations for controlled model geometry, X-ray fluorescence (XRF) spectra from colored glass rods, and laser-induced breakdown spectroscopy (LIBS) data from Dalbergia wood species. Results revealed that uncertainty increases with spectral similarity and perpendicular alignment between noise structures and decision boundaries. In real-world applications, classification metrics alone proved insufficient to assess model reliability. The inclusion of uncertainty intervals enabled identification of ambiguous predictions even in cases of perfect classification accuracy. This work advances chemometric analysis by linking measurement uncertainty to classification outcomes, offering a robust framework for decision-making in high-stakes analytical contexts.

The effectiveness of artificial neural networks, which were key technologies in artificial intelligence, greatly depends on the quality of the input data. Data cleaning, a crucial component of data preprocessing, played a vital role in enhancing the accuracy, robustness, and generalization capabilities of neural network models. In this study, a Feedback-Driven Iterative Cleaning (FDIC) framework, guided by model performance, was developed and applied to the study of droplet size prediction models for nanocellulose-stabilized Pickering emulsion systems. After randomly removing between 1% and 40% of the data, an artificial neural network model was established using CNC particle size (X1), CNC concentration (X2), and the oil–water volume ratio of CNC to oil-phase monomer (X3) as input variables, with emulsion droplet size (Y) as the quantitative index. The model's accuracy was evaluated after data removal using the coefficient of determination (R²), mean squared error (MSE), and mean absolute scaling error (MASE). The main finding was that targeted removal of a small portion of the data significantly improved the predictive power of the model. Specifically, removing 5% of the dataset results in optimal performance, with R² improving from 0.5307 without cleaning to 0.7258, with an MSE of 183.4917, and MASE of 0.4060. This result suggested a significant and quantifiable improvement in the accuracy of the model through our iterative cleaning process. The study revealed a nonlinear relationship between the number of iterations and the model's generalization ability. This finding offered a novel methodological tool for data governance in the smart era and demonstrates significant value in dynamic environments.

In order to achieve nondestructive analysis and identification of the relative formation time of paper evidence and to solve the difficulties in document authenticity identification in the field of forensic science, this study selected three-dimensional fluorescence spectroscopy data of paper evidence of the same brand and model collected in the same storage environment within the last decade (2012–2023). After preprocessing steps like eliminating scattering, smoothing noise and principal component analysis (PCA), machine learning algorithms such as K-nearest neighbor (KNN) and linear discriminant analysis (LDA) were employed to classify and predict specific feature bands. The accuracy of KNN and LDA was 94.5% and 98.9%, respectively. Furthermore, relative formation time prediction was conducted for paper samples by LDA in the sample library, achieving an accuracy rate of 98.0%. Finally, the established model was successfully applied to analyze an actual case involving suspected “forged official documents.” It accurately determined the relative formation time of the forged paper, and the analysis results were consistent with the suspect's confession.

Ion mobility spectrometry–mass spectrometry (IMS-MS) enables rapid acquisition of collision cross section (CCS), a critical physicochemical property for analyte characterization. Despite CCS being theoretically defined as the rotationally averaged projected area of 3D atomic spheres, existing models have underutilized this geometric insight. Here, we present a projected area–based CCS prediction method (PACCS). It integrates voxel-projected area approximation, graph neural network (GNN)–extracted features, and m/z to achieve accurate and rational CCS prediction. A voxel-based algorithm efficiently calculates molecular projected areas by leveraging Fibonacci grids sampling and discretizing 3D conformers into voxel grids. PACCS demonstrates exceptional performance, achieving a median relative error (MedRE) of 1.03% and a coefficient of determination (R²) of 0.994 on the test set. External test set against AllCCS2, GraphCCS, SigmaCCS, CCSbase, and DeepCCS highlights the superiority of PACCS, with 80.1% of predictions exhibiting < 3% error. Notably, PACCS exhibits broad applicability across diverse molecular types, including environmental contaminants (R² = 0.954–0.979) and structurally complex phycotoxins (R² = 0.961), highlighting the superiority of PACCS in robustness and versatility. Computational efficiency is enhanced via parallelization, enabling large-scale CCS database generation (e.g., 5.9 million entries for ChEMBL within 10 h). Ablation studies confirm the pivotal role of voxel-projected areas (Pearson correlation coefficients > 0.988), while stability analyses reveal minimal sensitivity to conformational variability (standard deviation of R² is 0.00003). PACCS provides an open-source, scalable solution for expanding CCS databases, advancing compound identification in metabolomics and environmental analysis.

Human immunodeficiency virus (HIV) has far-reaching impacts on global public health. Acquired immunodeficiency syndrome (AIDS) has caused millions of deaths globally, with thousands still getting infected. Therefore, developing HIV-1 integrase inhibitors is crucial for controlling AIDS by slowing virus replication and transmission. This study is grounded in the framework of deep reinforcement learning, aiming to de novo design inhibitors of HIV-1 integrase-Lens Epithelial-Derived Growth Factor/p75 interaction and subsequently employing molecular docking to screen potential therapeutic compounds. Initially, a molecular generation model was established based on the long short-term memory algorithm and refined through transfer learning to obtain a preliminary generative model. Subsequently, the deep reinforcement learning strategy was employed, using inhibition activity as a reward value, enabling the model more likely to generate molecules with desirable properties. The results indicate that the reinforced generation model not only generates novel and effective SMILES structures with medicinal potential but also demonstrates strong binding affinity between the generated molecules and the target protein, as indicated by molecular docking experiments. Ultimately, through virtual screening, we identified six lead compounds having the potential to become inhibitors of interaction between Lens Epithelial-Derived Growth Factor/p75 and HIV-1 integrase, providing an effective and practical strategy for de novo drug design of HIV-1 integrase inhibitors.