Group testing is widely used to improve screening efficiency in populations with low prevalence, but pooling may reduce assay sensitivity through dilution. Existing statistical work often incorporates this effect as a modeling assumption or design constraint, while the question of whether a dilution effect is present in a specific testing procedure is less often treated as a formal inferential problem. This paper develops a hypothesis testing framework for assessing dilution effects under a binary measurement error model. Dilution is represented by an increase in the group-specific false negative rate as pool size increases, with specificity assumed to remain stable across pool sizes. We derive closed-form estimators and one-sided test statistics for both unknown- and known-prevalence settings, using normal and Poisson approximations together with null-bootstrap critical-value calibration. Simulation studies show that the proposed procedures control type I error reasonably well and gain power as the dilution effect becomes stronger, although detection is more difficult under very low prevalence. Applications to infectious disease screening settings and SARS-CoV-2 pooled NAAT data illustrate how the framework can quantify evidence for reduced sensitivity under pooling.
Accurate monitoring of ground-level ozone and coexisting air pollutants remains challenging for low-cost metal oxide semiconductor (MOS) sensor arrays because their cross-sensitive responses are jointly governed by inter-sensor spatial correlations, gas-surface temporal dynamics, and task conflicts between gas identification and concentration regression. Here, we propose Spatiotemporal feature Decoupling with Dynamic task Coordination for Multi-Task Learning (STDC-MTL), an uncertainty-coordinated learning framework for simultaneous gas classification and concentration regression using MOS sensor-array signals. STDC-MTL explicitly integrates a hybrid spatiotemporal encoder for array-temporal correlation modeling, a dual-branch local-global feature learning module that separately captures neighborhood temporal variations and long-range response dependencies, and a Bayesian uncertainty-driven coordination mechanism for adaptive task balancing. Experimental results on a custom dataset of ozone, acetone, NO, NO2, and NH3 show that STDC-MTL achieves 100% classification accuracy and a concentration regression R2 of 0.998. More importantly, it reduces the mean absolute error (MAE) by 42.6% compared with the state-of-the-art Gaussian-TCN baseline, with an average normalized root mean square error (NRMSE) of 0.0138 and only 0.37 million parameters. These results demonstrate that the proposed framework improves quantitative prediction under cross-sensitive MOS responses while maintaining a lightweight architecture, providing an effective algorithmic strategy for intelligent air-pollutant sensing.
Fault detection in multivariate dynamic processes remains challenging because of strong variable correlations, temporal dependencies, and limited abnormal labels. Conventional statistical monitoring may be insufficient for weak dynamic deviations, while deep learning models do not always exploit compact latent representations or explicit temporal information. To address these issues, this study proposes a latent predictive monitoring framework that combines principal component analysis (PCA), learnable temporal encoding, and gated recurrent unit (GRU)-based multi-step prediction. Process measurements are projected into a compact PCA score space, augmented with explicit temporal representations, and monitored through prediction residuals between observed and predicted latent trajectories. The proposed framework is evaluated on the Benchmark Simulation Model No. 1 and further validated using industrial papermaking process data. On the BSM1 benchmark, the method achieved an average fault detection rate of 88.02 ± 2.73% with an average false alarm rate of 1.62 ± 0.15%, outperforming classical statistical methods and several deep learning baselines, particularly in slowly evolving and weakly distinguishable fault scenarios. The industrial case further demonstrated the transferability of the framework while revealing section-dependent false alarm behavior. These results indicate that integrating latent space projection with explicit temporal representation can improve the separability of prediction residuals in multivariate dynamic monitoring tasks.
Chemometrics has become an essential tool in analytical chemistry due to the increasing complexity and dimensionality of modern datasets. However, many researchers and students lack both the coding skills and a clear methodological roadmap required to perform a comprehensive chemometric analysis. In this work, we propose a structured, question-driven workflow comprising 17 sequential questions that guide users through a complete chemometric analysis using large language models (LLMs). The workflow includes data inspection, assessment of normality, identification of discriminative variables, unsupervised analyses by Principal Component Analysis (PCA) and Principal Coordinates Analysis (PCoA), and construction and validation of Partial Least Squares Discriminant Analysis (PLS-DA) models. The workflow was demonstrated using a previously reported lung cancer lipidomics dataset (Sci. Transl. Med., 14, eabk2756, 2022) containing the concentrations of nine lipids measured in healthy controls and lung cancer patients. All analyses were performed in R using scripts generated by ChatGPT (GPT-5.5), Google Gemini (Flash-Lite), and DeepSeek (V4) in response to each analytical question. The generated scripts were executed directly in R to produce all figures and statistical analyses. The three LLMs successfully reproduced the complete exploratory chemometric workflow, whereas reproducible PLS-DA validation required explicit specification of preprocessing and cross-validation parameters. These results demonstrate that LLMs can substantially lower the technical barrier to chemometric analysis by providing methodological guidance and executable R code, making rigorous multivariate data analysis more accessible for both education and research.
Leishmaniasis remains a major neglected tropical disease, and the discovery of new therapeutic options is hindered by high costs, long development timelines, and the parasite's biological complexity. In this context, artificial intelligence has emerged as a promising strategy to accelerate and refine different stages of the drug discovery process. This scoping review aimed to map how artificial intelligence has been applied in the discovery of new treatments for leishmaniasis, identifying the main computational approaches, their applications, the validation strategies employed (including cross-validation and external testing), and the translational outputs reported in the literature. The review was conducted in accordance with the Joanna Briggs Institute recommendations, and studies were retrieved from PubMed, Scopus, and Web of Science using predefined eligibility criteria. The data report was prepared in accordance with the PRISMA-ScR checklist. The included studies showed that artificial intelligence has been applied across multiple stages of antileishmanial drug discovery, including target identification and validation, ligand-based and structure-based virtual screening, hit discovery, lead optimization, and pharmacokinetic and toxicity prediction. Random Forest was the most frequently reported model, particularly in ligand-based virtual screening and quantitative structure–activity relationship modeling, whereas Support Vector Machine, Decision Tree, Artificial Neural Networks, clustering approaches, and deep learning tools such as AlphaFold, DeepLoc, DeepPurpose, and MONN were also identified. The most common applications involved structural relationship prediction, especially binary structure–activity relationship classification, followed by pattern recognition and screening, biological identification and characterization, and descriptor selection. Overall, the findings indicate that artificial intelligence has become a versatile and increasingly relevant tool in antileishmanial drug discovery. However, important limitations remain, including data scarcity, dataset imbalance (32.7% of the included studies did not report applying to a class-imbalance-handling procedure), a lack of parasite-specific benchmarks, and the predominance of studies restricted to in silico validation. Even so, integrating artificial intelligence with experimental validation may substantially strengthen future efforts to identify new treatments for leishmaniasis.
Accurate interpretation of urine test strips is vital for point-of-care (POC) diagnostics; however, manual readings are constrained by subjectivity, inter-observer variability, and inconsistent environmental conditions. This study presents a systematic comparison of three segmentation paradigms, K-means clustering, Density-Based Spatial Clustering of Applications with Noise (DBSCAN), and the Segment Anything Model (SAM), for automated identification and delineation of reagent pads on ACON MISSION 10-parameter urine test strips. A dedicated portable imaging system was developed using a Raspberry Pi 4 with controlled LED-based illumination inside a fully enclosed 3D-printed chamber. Analytically prepared synthetic urine solutions (S1–S4) spanning clinically relevant concentration ranges for seven analytes, including glucose (500–2000mg/dL), protein (50–200mg/dL), and occult blood (0.75–3.0mg/dL), were used to generate a dataset of 120 images across five test conditions. All experiments were performed with five replicates per condition and five image captures per strip to assess reproducibility. From aggregate confusion-matrix analysis, SAM achieved 99.34% pixel-wise accuracy (± 0.15%), 94.34% precision, 87.71% recall, and an F1-score of 0.909, compared to K-means (96.57% accuracy, 52.69% precision, F1 = 0.622) and DBSCAN (94.56% accuracy, 38.46% precision, F1 = 0.513). The mean Intersection over Union (IoU) of 83.32% (± 0.54%) further confirms SAM’s superior ability to delineate reagent pad boundaries with high fidelity. Statistical analysis revealed that SAM maintained a coefficient of variation below 0.5% across all concentration levels, demonstrating robust invariance to chromatic shifts. While this proof-of-concept utilizes synthetic samples under controlled conditions, the results demonstrate the feasibility of deploying transformer-based segmentation on low-cost embedded hardware for objective, automated urinalysis.
DNA-binding protein prediction helps to understand the fundamental process of life activities. Existing computational methods targeting large-scale DBPs are cost-effective and fast compared to conventional biochemical experiments. However, many earlier in-silico techniques rely on shallow or handcrafted descriptors that may not fully capture contextual sequence information. Recent pre-trained biological language models (BLMs) improve sequence representation by learning rich contextual characteristics from proteins. Methods In this study, we develop CLKAN-DBPred, an integrated framework for predicting binding activity of DBPs using a fine-tuned approach that integrates DistilProtBert, Prot-T5, and ESM2 embeddings, the RFE-F7-1195 representation, and a convolutional KAN classifier. Training used hard-label cross-entropy together with a KL-divergence term computed against smoothed targets (epsilon = 0.10; kl_weight = 1.5). Additional component-wise loss ablations, KAN pathway and spline-parameter sensitivity analyses, pure machine-learning baselines, and paired statistical tests were conducted. Results CLKAN-DBPred achieved 85.04 Acc, 0.7009 MCC, 0.8489 F1, and 0.9075 AUC in the reported fivefold CV evaluation and 81.76 Acc, 0.6352 MCC, 0.8163 F1, and 0.8905 AUC on PDB296. It obtained the highest AUC and MCC among the compared DBP predictors. The added analyses localized most of the KAN contribution to the spline pathway and showed limited sensitivity to the tested loss and spline settings. Against a direct XGBoost baseline on RFE-F7-1195, CLKAN-DBPred achieved stronger mean CV performance; paired independent-test differences in Acc, MCC, F1, and AUC were not significant. These results support CLKAN-DBPred as a competitive sequence-based framework for DBP prediction. All data and models are available at: https://doi.org/10.5281/zenodo.18023779
In modern industrial processes, the accurate detection of key performance indicators (KPIs) is often challenged by high system nonlinearity compounded by incipient anomaly characteristics. To address this issue, a hybrid kernel deep orthogonal subspace analysis algorithm is proposed in this paper. First, a hybrid kernel combining Gaussian and polynomial kernels is utilized to project the original nonlinear process data onto a high-dimensional feature space. This projection enhances linear separability, while the Gaussian component preserves local neighborhood similarities and the polynomial component captures global high-order feature interactions. Within this feature space, a kernel correlation matrix (KCM) is constructed to characterize the associative properties of the relevant information between process inputs and KPIs. A deep singular value decomposition scheme is then applied to this KCM to hierarchically decompose the KPI-related information into multiple kernel subspaces. Hierarchical decomposition further enables the construction of a multi-order subspace structure, which significantly enhances sensitivity to incipient anomalies that are easily obscured by noise in first-order projections. Furthermore, complexity is assessed via floating-point operations, serving to illustrate the feasibility for both the training and detection phases. Finally, experiments on the penicillin fermentation process and thermal power plant process are carried out. The practical applicability of the proposed method is verified, and superior performance over state-of-the-art techniques is demonstrated.
Basic oxygen furnace (BOF) steelmaking is a high-temperature chemical reaction process where endpoint carbon content and temperature are critical variables determining molten steel quality. Traditional data-driven models rely on manual feature engineering and iterative hyperparameter tuning, with predictions difficult to translate into actionable guidelines. This paper proposes an interpretable soft sensor framework integrating large language model (LLM)-based knowledge feature engineering, Tabular Prior-data Fitted Networks (TabPFN), and SHapley Additive exPlanations (SHAP)-based threshold analysis. The framework generates physicochemical knowledge-enhanced features via multiple LLMs, combines correlation screening with recursive feature elimination with cross-validation (RFECV) to select the optimal feature subset, applies TabPFN for rapid endpoint prediction, and extracts sensitive intervals and operational thresholds via SHAP dependence fitting. Using 814 heats from a 220-ton BOF, the model achieved test-set coefficient of determination (R2) of 0.5318 and 0.8315, hit rates (HR) of 79.75% (±0.01%) and 92.64% (±10 °C), and training times of 0.3805 s and 0.3944 s for carbon content and temperature, respectively. Against state-of-the-art models, R2 increased by 5.75% and 1.66%, HR by 3.06 and 1.23 percentage points, and training speeds by factors of 402 and 453, respectively. SHAP analysis identified a high-sensitivity interval of [1682, 1792] kg for light-fired dolomite on carbon content, and an efficient temperature regulation region within 700 kg of ore consumption. Over 52 industrial heats, HR reached 94.23% and 100% for carbon content and temperature. These results demonstrate that the framework enables rapid BOF endpoint prediction and operational threshold extraction, providing a quantitative basis for endpoint control.
Food adulteration represents a significant challenge to ensuring authenticity, particularly in powdered matrices such as corn flour. In this study, a rapid and non-destructive methodology was developed to detect corn flour adulteration with black soldier fly (Hermetia illucens) flour using portable near-infrared (NIR) spectroscopy combined with DD-SIMCA modelling and LOVE variable selection. Authentic and adulterated samples were analysed at adulteration levels ranging from 1% to 20%, and spectra were acquired in the 900-1600 nm range. Spectral data were pre-processed using the Savitzky–Golay first derivative, standard normal variate (SNV), and mean centering. PCA revealed a compositional gradient associated with increasing adulteration level, although some overlap remained at the lowest concentrations. Application of the LOVE method reduced the number of variables from 212 to 92 while retaining the most informative spectral regions for model development. The DD-SIMCA model developed using the selected variables outperformed the full-spectrum model, achieving a sensitivity of 1.00 for the target class and a specificity of 1.00 in the independent external test set. The results demonstrate the potential of the proposed methodology for food authenticity assessment under the experimental conditions investigated.
To enhance the prediction accuracy of key variables in penicillin fermentation under small-sample constraints, this paper presents a transfer learning framework utilizing Long Short-Term Memory (LSTM) networks with L2-distance to Starting Point (L2-SP) regularization. A deep temporal feature extractor is pre-trained on source-domain data to capture general fermentation dynamics and long-range dependencies. The learned representations are subsequently transferred to the target domain as a parameter initialization, enabling knowledge reuse under data-scarce conditions. A dedicated regularization strategy is introduced during fine-tuning to constrain the deviation of model parameters from their initial pre-trained values, thereby preserving transferable knowledge while preventing overfitting on limited target samples. The simulation results demonstrate that the proposed framework generally outperforms training from scratch and conventional transfer learning in moderate-to-high sparsity regimes, while exhibiting decisive advantages over feature-alignment baselines such as CORAL. These findings validate the effectiveness of the proposed method and clarify its conditional applicability across varying data sparsity levels.
Accurate molecular toxicity prediction is essential for chemical safety assessment. Chemical language models (CLMs) offer powerful molecular representations, but their GPU-intensive fine-tuning limits accessibility. We present FewTox, a resource-efficient, CPU-trainable framework that combines frozen multi-source CLM features with multi-task learning for 12-endpoint toxicity prediction on Tox21. Under strict inductive evaluation, a compact MLP built on concatenated PubChem and ZINC ChemBERTa embeddings achieves a mean scaffold AUROC of 0.702. This performance approaches a fully fine-tuned ChemBERTa baseline (0.733) at approximately 160× lower training cost.ECFP-Random Forest remains a strong baseline (AUROC 0.728), demonstrating the effectiveness of explicit substructure fingerprints under scaffold splitting. FewTox complements fingerprint-based models by integrating ECFP substructures with continuous CLM representations; this hybrid reaches an AUROC of 0.725. The multi-task neural head also supports post-hoc temperature scaling, which improves NLL and Brier score without changing discrimination. On the two-endpoint ClinTox benchmark, FewTox achieves a scaffold AUROC of 0.979. These results establish a practical accuracy–efficiency profile for frozen molecular representations in resource-constrained toxicity modeling.
Generative molecular models produce candidate libraries containing structures with poor descriptor profiles or structural alerts that can be removed before docking or expert review. We test whether this early selection can be implemented as a reproducible filtering and ranking protocol. LiteGov applies medicinal-chemistry and structural-alert filters and ranks retained molecules using the quantitative estimate of drug-likeness (QED) and synthetic accessibility (SA), with fixed 0.8/0.2 weights. Across evaluations, mean QED increases from 0.665 to 0.887, mean SA decreases from 2.818 to 2.198, and pan-assay interference compound (PAINS) alerts decrease from 4.6% to 0%. In an exact ablation of 10,678 unique structures, LiteGov attains Spearman correlation 0.911, Top-100 overlap 0.620, and normalized discounted cumulative gain 0.977 against an author-specified property-priority reference; relative to QED-only ranking, Top-100 SA decreases by 0.232 with a QED difference of −0.003. Generator-held-out calibration selects weights of 0.64–0.68 and raises mean property-reference Spearman correlation from 0.900 to 0.917. AiZynthFinder route-solution rates increase from 0.58 to 1.00 for GenMol and from 0.64 to 0.80 for MoLeR. Complementary structure-aware audits report Spearman correlation 0.067 and Top-100 overlap 0.028 with an AutoDock Vina–GNINA affinity consensus, alongside a ChEMBL activity-label area under the receiver operating characteristic curve of 0.369. These evidence layers establish LiteGov as a descriptor-based triage and computational route-feasibility protocol for preparing candidate libraries for target-aware screening.
Accurate modeling of temporal dependencies is crucial for industrial soft sensing of time series, where deep learning approaches have achieved remarkable progress. However, time series collected from real industrial processes are often irregularly sampled due to practical factors such as sensor malfunctions, communication interruptions, and data corruption. The irregularly sampled time series poses significant challenges for most deep learning models, including recurrent neural networks and Transformers, as traditional dynamic soft sensing methods cannot directly handle heterogeneous sampling intervals. Addressing this issue, this article proposes a time-aligned predictive learning (TAPL) framework that explicitly models sampling time intervals to effectively capture temporal dependencies among irregularly sampled sequences. TAPL consists of two core modules, including a Fourier position embedding-based time-aware attention and a positional time-aligned recurrent network. The former integrates sampling time information into attention score computation and assigns adaptive weights to modulate temporal correlations, while the latter incorporates time intervals into recurrent updates, adjusting the influence of past observations on current states. Experiments on real-world industrial datasets verify the efficacy and superiority of the proposed approach.
