The rapid global expansion of β-lactamase-mediated antimicrobial resistance demands mechanistic approaches capable of resolving the dynamics of enzyme adaptation. Although β-lactamase evolution often involves subtle rearrangements rather than large structural shifts, traditional structural and simulation analyses struggle to capture the conformational heterogeneity that underlies shifts in substrate specificity and inhibitor susceptibility. Here, we review recent advances in applying deep learning to probe the conformational dynamics of β-lactamases across classes A-D. We highlight how convolutional variational autoencoders (CVAEs) reconstruct nonlinear conformational manifolds from molecular dynamics simulations, exposing metastable states, cryptic pockets, and catalytic intermediates. DiffNets integrate supervised objectives to identify structural determinants of biochemical phenotypes, while BindSiteS-CNN and geometric deep learning methods provide high-resolution insight into active-site remodelling and local pocket plasticity. Additionally, graph neural networks trained on dynamics-informed descriptors capture long-range allosteric couplings and accurately predict mutational fitness and epistasis. The deep learning-enabled analysis of protein dynamics offers a unified and predictive framework for understanding β-lactamase adaptation.

Accurate online detection or prediction of key quality variables provides critical reference information for optimizing and controlling operating variables in industrial processes. However, frequent fluctuations in raw material properties and environmental conditions often give rise to multiple data distribution modes within the same production process. Moreover, the inherent uncertainties and the energy-material coupling characteristics of industrial processes make it particularly challenging to uncover the underlying topological relationships among process variables. To address these issues, this article proposes a novel jointly shared-specific variational graph attention autoencoder (JSS-VGATE) model for spatial topological feature extraction and key quality variable prediction in multimode industrial processes. Specifically, a variational graph attention autoencoder is first constructed, which combines graph attention mechanisms with the variational inference architecture to adaptively learn the dynamic correlation strengths between adjacent nodes, thereby capturing complex variable interactions. Subsequently, a comprehensive loss function is designed to achieve high-fidelity extraction of representative latent feature distributions. Furthermore, a cross-mode jointly shared-specific learning framework is developed to simultaneously capture global shared features across modalities and preserve local specific features of each modality, while a learnable gated fusion mechanism is introduced to balance modality invariance and heterogeneity, thereby enhancing cross-mode information integration. Finally, the effectiveness and superiority of the proposed JSS-VGATE are validated on two representative real-world industrial datasets compared to other state-of-the-art methods.

A soft sensor regression modeling method for complex chemical processes based on Variational Autoencoders and vine copula

Wire-laser directed energy deposition (WL-DED) enables the fabrication of large-scale metallic components but frequently suffers from process-induced surface defects that hinder part quality and increase post-processing costs. Automated inspection is challenging because defects are diverse and rare, making large labelled datasets impractical. This paper proposes an interpretable semi-supervised framework for surface detection and localization on WL-DED components using high-density 3D point clouds acquired by laser scanning. The workflow includes point-cloud preprocessing, patch-based segmentation, voxelization, and semi-supervised representation learning of defect-free surface morphology. Two 3D deep autoencoder models, i.e., a convolutional autoencoder (CAE) and a variational autoencoder (VAE), are trained exclusively on normal patches and detect anomalies through voxel-wise reconstruction errors. Defects are localized by mapping reconstruction-error heatmaps back onto the original surface, enabling quantitative visualization of defect severity. Experimental results on WL-DED thin-wall samples show that the optimized CAE achieves 86.09% precision, while the VAE reaches 86.43% precision with improved defect localization (mIoU up to 0.7234). Activation-map analysis provides interpretability by highlighting geometric regions that drive anomaly responses. A hyperparameter study demonstrates that lower voxel resolutions and smaller patch sizes improve robustness and reduce false positives. The proposed framework generalizes to more complex multi-bead, multi-layer structures with minimal retraining, supporting practical deployment for intelligent inspection and decision-making in additive manufacturing quality assurance.

A fault diagnosis method based on diffusion model and 2D-CNN for small sample conditions: Application to reciprocating pumps

MCVae-CFM: a behavior-embedded dual calibration framework for time-varying car-following models via multichannel variational autoencoding

Physics-informed sequential data augmentation for three-phase flow modeling

3DDF-VAE: Dual-frequency variational autoencoder with pose-consistency validation for rare cryo-EM conformation discovery.

PatientFlow: Learning to generate mixed-type longitudinal clinical data with flow matching.

High-performance dielectric materials are strategically vital for post-Moore microelectronics, high-power electrostatic capacitors, and flexible advanced electronics, yet their design is constrained by the intrinsic trade-off between breakdown strength (Eb) and relative permittivity (εr). Traditional high-throughput screening is limited to existing chemical spaces and cannot resolve the inherent orthogonality between key dielectric metrics and multiphysics coupling conflicts. This review systematically maps the paradigm shift in dielectric research from passive empirical screening toward generative, AI-driven autonomous discovery. We first examine the multiscale physical origins of the core dielectric performance trade-offs and the challenge of high-fidelity data scarcity, highlighting the role of physics-informed descriptors in bridging atomic-scale structures and macroscopic properties. We then survey the evolution of high-fidelity surrogate models and multiobjective optimization frameworks for efficient structure-property mapping followed by analysis of mainstream generative architectures (variational autoencoders, VAEs; generative adversarial networks, GANs; diffusion models), which enable the inverse design of dielectric polymers and inorganic crystals beyond known chemistries. Furthermore, we discuss the integration of self-driving autonomous laboratories and active learning strategies to close the feedback loop between computational prediction and experimental validation. Finally, we address the unresolved barriers in physics embedding, model explainability, and experimental synthesizability, outlining an actionable roadmap toward Physics-Sovereign AI and large language model (LLM)-driven digital scientists for dielectric innovation. This review provides a holistic, application-focused perspective on the transition from Edisonian trial-and-error approaches to a new era of rational, accelerated dielectric materials development and delivers a practical decision-making framework for researchers selecting AI tools for targeted dielectric design.

The head-up tilt table test (HUTT) is a lengthy and uncomfortable procedure for patients which often induces fainting. Post-transient loss of consciousness, nausea, vomiting, and pallor may occur. This study aims to develop and evaluate the efficacy of anomaly detection methods based on autoencoding for early prediction of syncope onset, enabling preemptive HUTT termination and thereby avoiding unnecessary discomfort. The four input signals: heart rate (HR), systolic blood pressure (SBP), diastolic blood pressure (DBP), and high frequency normalized RRI (Hfnu_RRI) were processed into feature images for explicit correlation encoding. The feature images served as input for the syncope detector, which consisted of an autoencoder (AE) coupled with an external algorithm that computed the severity of the cardiac anomaly. The choice of AE to model test-negative patients was evaluated across six unique candidate architectures. The best architecture was fine-tuned with the Keras Bayesian tuner. Lastly, the overall syncope detector was validated with 100 iterations. The developed temporal convolutional autoencoder anomaly detector (TCAAD) attained an accuracy of 0.9424, a recall of 0.9838, a precision of 0.9141, a specificity of 0.9009, an F1 score of 0.9461, and an early prediction time of 523.69s. The model's performance was comparable to other real-time prediction methods evaluated in this study, and demonstrated one of the longest early prediction times. This highlights the effectiveness of anomaly detection methods combined with signal correlation monitoring.

Chronic Obstructive Pulmonary Disease (COPD) is a complex, heterogeneous disease. Traditional subtyping methods generally focus on either the clinical manifestations or the molecular endotypes of the disease, leading to classifications that only partially reflect disease heterogeneity. Here, we introduce a variational autoencoder-based subtyping pipeline that jointly embeds clinical and gene expression data into a single subject-level representation. We evaluate the framework in the COPDGene study, a large study of current and former smoking individuals with and without COPD. Prediction experiments show that the embeddings have predictive accuracy comparable to or better than other unsupervised embedding approaches. Using trajectory learning approaches, we identify five well-separated subtypes with distinct clinical phenotypes, expression signatures, and longitudinal outcomes. Finally, we show that our findings generalize to an external validation cohort. Overall, our approach enables a transition from isolated phenotypic or molecular subtyping toward an integrated and clinically meaningful understanding of COPD heterogeneity.

Deep learning (DL)-based pathological image modelling and analysis approaches offer transformative potential for early cancer diagnostics, yet limited sample sizes and a lack of interpretability often hinder efficient clinical translation. Here, we present the interpretable Multi-Task Digital Pathology Model (iMDPath), an end-to-end, highly explainable multi-task deep learning framework that simultaneously addresses these challenges by integrating data augmentation, diagnostic prediction, and visualization of pathological image features. The iMDPath comprises three modules: Augmentation (iMDPath-Aug), Prediction (iMDPath-Pred), and Visualization (iMDPath-Vis). iMDPath-Aug incorporates a vector-quantized variational autoencoder (VQ-VAE) for enhanced data augmentation, capturing essential pathological features from limited datasets. A Swin Transformer-Based (Swin-B) predictor in the iMDPath-Pred module leverages the augmented data to achieve better performance than patch-level and foundation-model-based encoders such as InceptionV3 and Phikon across six diverse cancer pathology datasets, including gastric, breast, lung, and colorectal cancer. Finally, iMDPath-Vis, a novel visualization module combining the full gradient (FullGrad) and occlusion sensitivity analysis, provides pathologists with actionable insights by highlighting the specific tissue regions driving model predictions. Overall, iMDPath not only surpasses existing methods in diagnostic accuracy, sensitivity, and generalization across these datasets, but also offers a transparent and interpretable AI solution for precision oncology, paving the way for more reliable and efficient clinical decision-making.

Accurate fault pattern recognition in the flotation process is crucial for rapid fault response and reducing production risks. However, dynamic flotation process data exhibit high dimensionality, nonlinearity, and nonstationarity, along with significant noise and uncertainty, which severely impact the accuracy (ACC) of fault pattern recognition. To address these challenges, we propose a fault pattern recognition method for froth flotation based on higher order spatial-temporal block and dual-stream variational graph neural networks (HoStB-DVGNNs). First, we construct a high-order spatiotemporal block using the groups of key frame images to comprehensively characterize the flotation production conditions. Then, we develop a dual-stream variational graph neural network (DVGNN) that includes an apparent feature stream and HoStB stream to effectively extract dynamic froth information. In addition, we introduce a bilateral self-supervision mechanism to build a variational autoencoder (VAE), significantly enhancing the model's generalization performance. Finally, extensive experiments on benchmark datasets and real flotation processes validate the effectiveness and robustness of the proposed fault pattern recognition method.

Objective.Quantitative analysis of dynamic positron emission tomography (PET) scans requires knowledge of the arterial input function (AIF). Existing means of extracting the AIF are invasive and costly (blood sampling), or come with significant errors (image-derived input function, IDIF). We present a novel image-derived AIF method using a machine learning technique that does not require external training data.Approach.Voxel-by-voxel time-activity curves are used as individual input samples for training a customised autoencoder machine learning model. Autoencoders (AEs) are models that map input samples to themselves, with an intermediate latent layer with few nodes. This drives the training algorithm to find an optimal bottleneck representation of the input. The IDIF is extracted from the weights of the trained model.Main Results.The method was evaluated on dynamic PET scans of rats with TSPO tracer [18F]LW223. Volumes of distribution (VT) from arterial blood sampling (ground truth) were compared using Logan plots with IDIF-AE (mean absolute percentage error ±32%) and conventional IDIF from left ventricle (±54%). The method was also successfully adapted for scans of mice with neuro PET tracer [18F]SynVesT-1.Significance.This study demonstrates a novel machine-learning based image-derived input function for dynamic PET that can outperform classical IDIFs in determining kinetic parameterVT, without requiring external training data.

ABSTRACT Using the potential spatiotemporal features in trajectory data to identify the operation mode of agricultural machinery is an important basic task in the field of precision agriculture. However, existing methods for identifying the operation mode of agricultural machinery trajectories fail to model the dependencies of agricultural machinery trajectories from different ranges, and the imbalanced data distributions of different categories of agricultural machinery trajectory data produce identification bias. To overcome the above defects, this paper proposes a multi-range spatiotemporal information capture network based on amplified feature deformation (VRPNet). First, to solve the identification bias caused by the imbalanced data distribution of the model, we design a data balancing module based on a factorized variational autoencoder (FVAE), which independently factors and encodes the features of minority class trajectory samples and then decodes and generates quasi-trajectories similar to the original trajectory points to balance the data distributions of different categories. Second, to explore the potential spatiotemporal information of trajectories fully, we propose a trajectory information multi-scale amplification module, which applies kinematic methods and statistical methods to extract multi-scale features of agricultural machinery trajectories in different spatiotemporal ranges to mine the inherent information of agricultural machinery trajectories. Finally, to comprehensively model the dependencies of agricultural machinery trajectories, we propose a spatial correlation capture module based on a low-rank approximation matrix (LRSC) and a dynamic multi-path convolution bottleneck based on feature deformation (FD-DPC) to assemble into a trajectory context encoder to explore the feature interactions of agricultural machinery trajectories in different ranges. To verify the effectiveness of the method, we conducted experiments on paddy and wheat trajectory public datasets. The results show that the accuracies of VRPNet on the paddy and wheat trajectory datasets are 94.20% and 96.06%, respectively, which are improvements of 1.72% and 1.92%, respectively, over those of the currently widely used models.

In the analysis of diverse omics data, a common and important preliminary step involves computing low-dimensional embeddings using techniques such as PCA, UMAP, t-SNE, or variational autoencoders. These embeddings provide a global overview of sample distributions and their relationships, often serving as the basis for formulating biological hypotheses. To facilitate rapid and intuitive exploration of such low-dimensional embeddings, we developed Yomix, an interactive omics-agnostic visualization and data exploration tool. Yomix enables users to flexibly define subsets of interest using a lasso selection tool, instantly compute their feature signatures, and compare their distributions. Yomix is a fast and efficient tool for interactive exploration of diverse omics datasets. Yomix and its documentation are publicly available at https://github.com/perrin-isir/yomix. Supplementary data are available at Bioinformatics online.

Polymer nanocomposites exhibit highly nonlinear viscoelastic behavior influenced by complex multiscale microstructural interactions between nanofillers and the polymer matrix. Existing data-driven models such as ANN, RNN, GRU, and LightGBM show limitations in capturing spatial-temporal operator dynamics, interfacial energetics, and frequency-dependent mechanical responses. These models often rely on local approximations, struggle with generalization beyond training distributions, and lack the ability to generate optimized microstructures. To address these limitations, this study develops a Hierarchical Neural Operator-Based Multiscale Learning Framework for accurate nonlinear viscoelastic response prediction and microstructure optimization in polymer nanocomposites. The proposed framework aims to establish a physics-aware data-driven pipeline capable of predicting stress-strain behavior, storage modulus G' (Pa), relaxation times τ (s), and glass-transition shift ΔTg​, while simultaneously identifying optimal microstructures. The methodology integrates a Fourier Neural Operator (FNO) surrogate for learning continuous mechanical response operators, a Convolutional Variational Autoencoder (CVAE) for generative microstructure latent representation, and a Physics-Guided Multi-Stage Atom Search Optimization (PG-MS-ASO) module for constrained microstructure design. Python, NumPy, scikit-learn, PyTorch, and Matplotlib form the computational toolset. Key findings indicate strong predictive accuracy, with the hierarchical model achieving RMSE = 0.15, MAE = 0.10, and R² = 0.98, outperforming RNN (R² = 0.92), ANN (R² = 0.91), GRU (R² = 0.97), and SNS-LightGBM (R² = 0.97). Optimized microstructures generated through CVAE + PG-MS-ASO exhibit significant improvements in performance metrics, with ΔTg reaching up to 31°C and storage modulus (G') exceeding 1.09 × 106Pa. Overall, the framework provides a robust pathway for linking microstructure design with macroscopic viscoelastic behavior, enabling accelerated materials discovery and offering a more generalizable and physics-consistent alternative to conventional machine-learning approaches.

The replicative senescence during in vitro expansion severely limits the clinical application of human umbilical cord mesenchymal stem cells (hUC-MSCs). Existing senescence assessment methods still face significant limitations in terms of noninvasive and real-time quantification. In this study, using atomic force microscopy (AFM), we systematically characterized the nanomorphology and mechanical properties of naive and senescent hUC-MSCs. The results revealed that senescent cells exhibited significantly increased height, surface roughness, adhesion, and elastic modulus compared to naive cells, along with enhanced bundling and formation of F-actin stress fibers. These findings show a new "senescence-associated mechanical phenotype" unique to hUC-MSCs. Notably, the hypoxic intervention effectively reversed these senescence-related mechanical changes, demonstrating the high sensitivity of AFM in detecting senescence. To achieve precise quantitative assessment of cell senescence, we also developed a deep learning model based on a variational autoencoder (VAE), which successfully established continuous low-dimensional representations of the age-related mechanical phenotype. This work exhibited excellent predictive performance and generalization ability under different culture conditions, enabling accurate prediction and early identification of hUC-MSCs senescence.

Single-cell RNA sequencing (scRNA-Seq) enables analysis of gene expression at single-cell resolution. RNA velocity analysis infers the temporal dynamics of transcriptional states from the relative abundances of spliced/unspliced mRNA quantified via scRNA-Seq.Classical RNA velocity approaches, such as scVelo, implement gene-specific kinetic modeling. Deep learning methods including DeepVelo, VeloVI, LatentVelo, SymVelo, and scTour are based on variational autoencoders (VAEs), which allow to enhance the robustness and accuracy by leveraging nonlinear latent representations. Here, we systematically evaluated the performance of deep learning RNA velocity tools by comparing with the scVelo dynamical model to access the possible advantages of VAE-base methods. For this purpose, public datasets (GSE149689 and GSE203233) were initially processed using a standard scRNA-Seq pipeline. Comparisons among results of selected velocity tools were conducted using cosine similarity of velocity vectors to assess directional concordance, and by mean squared error analysis of trajectory continuity for the deep learning models. Overall, VAE methods produced significant, richer, and more directionally coherent and consistent velocity fields than the classical model. Our findings indicate that deep learning models provide more consistent and biologically plausible cell-state trajectories, although at the expense of higher computational demands and reliance on accurate splicing quantification. Altogether, our results underscore the relevance of VAE-based frameworks to advance RNA velocity analysis while highlighting the need for careful preprocessing.