With the nascent reliance on unmanned aerial systems as a means of surveillance, security, and monitoring, it is essential to ensure that video data legitimacy is assured. Object-based Copy-Move Forgery (CMF), in which an object is either copied or moved in the frame or across a frame to alter the visual description, is one of the major threats to video integrity. Such manipulations are a problem for traditional detection methods because object motion, occlusion, and irregular background are complex. This research presents an interpretable and robust deep learning architecture to identify such forgeries, that is, a combination of transfer learning with a new Multi-Aspect Fossa Graph Transformer enhanced with Shuffle Attention (MAFGTN-SA). To achieve this, a better generative adversarial network, which is built upon GoogLeNet, is used to extract deep spatial-semantic features using normalized grayscale video frames. These characteristics are then represented by MAFGTN-SA to represent complex object relationships between the frames and spatial inconsistencies to accurately classify genuine and tampered frames. In order to increase interpretability, the model incorporates Local Interpretable Model-Agnostic Explanations (LIME) for saliency-based visualization of tampered regions. Experimental evaluations on three benchmark datasets, SULFA, GRIP, VTD, CG-1050 v2.0, and COVERAGE, validate the effectiveness of the proposed method, attaining an accuracy of 98.73%, 96.78%, 98.45%, 97.92%, and 97.36%, respectively, and with high F1-scores of 98.73%, 98.5%, 97.98%, 97.41%, and 96.93%. The results highlight the framework's superior detection capability, reliability, and explainability across diverse video forgery scenarios.

Heart failure (HF) is a common and life-threatening cardiovascular disease, and early accurate diagnosis is critical for improving patient survival rates and optimizing treatment outcomes. This work integrated the electrocardiogram (ECG) signals of patients with clinical text data to construct a 12-lead ECG-Text-LVEF Cardio dataset. A multimodal HF prediction model based on graph convolutional neural network (GCN) was proposed in this work. The model employs global feature vectors extracted from patients' ECG spectrograms as node features, generates medically semantic vectors by encoding clinical text via ClinicalBERT, establishes connections between nodes that meet a preset threshold after calculating cosine similarity between these vectors, and thus constructs a semantic adjacency graph that quantifies similarities in clinical descriptions among patients and explicitly models common associations in their symptom presentations and disease characteristics. Finally, it achieves reasoning from individual features to group relationships through 2-layer graph modeling. At the same time, the Top-K adjacency strategy and similarity threshold control the sparsity of edge connections, and the edge perturbation mechanism is introduced to enhance the generalization ability of the model. Experimental results demonstrate that the model achieves an accuracy of 98.5%, a precision of 98.4%, a recall of 99%, and an AUC of 0.997, validating the effectiveness of multimodal feature fusion in HF prediction and providing a practical pathway for applying GCNs to multimodal medical risk prediction tasks.

Early and objective screening of Autism Spectrum Disorder (ASD) remains challenging because conventional diagnosis primarily relies on behavioural assessment and clinical observation. To address this limitation, this study proposes a dual-domain computational framework for automated EEG-based ASD classification by integrating complementary time-frequency analysis with Horizontal Visibility Graph (HVG)-based network modelling. Four time-frequency decomposition techniques, namely the Short-Time Fourier Transform (STFT), Discrete Wavelet Transform (DWT), Wigner-Ville Distribution (WVD), and Superlet Transform (SLT), were employed to characterise the non-stationary dynamics of resting-state EEG signals. The resulting time-frequency representations were transformed into HVG networks, from which 17 graph-theoretic descriptors were extracted and evaluated using conventional machine learning classifiers, including a Soft Voting Ensemble. Among the investigated methods, the DWT-HVG framework combined with the Soft Voting Ensemble achieved the best performance, yielding an accuracy of 93.54%, sensitivity of 94.32%, specificity of 92.76%, F1-score of 93.62%, balanced accuracy of 93.54%, and an Area Under the Curve (AUC) of 98.17% using stratified 10-fold cross-validation. Statistical analysis using the Wilcoxon signed-rank test further confirmed the superiority of the DWT-based representation over the STFT, WVD, and SLT-based approaches. These findings demonstrate that integrating multiresolution time-frequency analysis with HVG-based graph-theoretic feature extraction provides an accurate, interpretable, and computationally efficient framework for objective EEG-based ASD screening.

Experimental determination of flash points (FPs) for liquid mixtures is laborious and costly, highlighting the need for reliable predictive approaches for safety assessment and engineering applications. Although numerous models have been reported for binary miscible mixtures, most rely on fixed model parameters or empirical correlations, which limits their ability to capture the nonlinear relationship between molecular structure and FP. In this study, a quantitative structure-property relationship (QSPR) framework that tightly integrates differential evolution (DE) with support vector regression (SVR) was developed to predict the FP values of binary miscible mixtures, where DE was employed to globally optimize key SVR hyperparameters and enhance model generalization capability. A dataset consisting of 332 compositions from 33 binary mixtures formed by pairwise combinations of 20 pure components was employed, and multiple molecular descriptor representation strategies were deliberately adopted to construct distinct DE-SVR models, enabling a systematic investigation of the combined effects of descriptor representation and model optimization on predictive performance. Three DE-SVR models were established based on different descriptor sets, and their predictive accuracy, robustness, and stability were comprehensively evaluated. Among them, the model constructed using physicochemical parameters exhibited the best overall performance. Comparative analyses with existing FP prediction methods reported in the literature further confirmed the effectiveness and superiority of the proposed DE-SVR-based models. The results of this study provide a practical tool for FP estimation of binary mixtures and offer valuable insights into the joint roles of model optimization and molecular representation in mixture property prediction.

Self-supervised learning and multimodal feature fusion approaches have led to significant advances in molecular property prediction. However, most existing multimodal fusion approaches rely on simple feature concatenation, which fails to effectively integrate multi-scale information from different molecular representations. To address this limitation, we propose a novel molecular representation learning framework, termed SMFP, which integrates self-supervised learning and multimodal feature fusion. The proposed framework leverages self-supervised learning to extract feature representations from SMILES sequences and molecular graphs, capturing both local and global molecular information from complementary perspectives, thereby improving the performance of downstream prediction tasks. Specifically, we design a multimodal pretraining framework that promotes feature fusion between SMILES sequences and molecular graphs. This framework incorporates a bidirectional cross-attention module for feature fusion, enabling the framework to better capture the relevance and importance of different molecular modalities and to more effectively integrate multimodal feature information. In addition, a non-overlapping adaptive masking strategy is introduced and applied to tokenized SMILES sequences and graph representations, further encouraging comprehensive feature fusion between the two modalities. Experimental results on eight classification datasets and six regression datasets from MoleculeNet demonstrate that the proposed method achieves competitive performance compared to representative baseline methods, showing advantages on 10 out of the 14 datasets, which validates its effectiveness.

Computational drug-target interaction (DTI) prediction provides a scalable alternative to costly and time-consuming experimental screening, but its reliability is limited by the scarcity of experimentally verified negative interactions. In public DTI databases, most unobserved drug-target pairs are unlabeled rather than true non-interactions. Randomly treating these unlabeled pairs as negatives can introduce label noise and reduce model reliability, particularly in cold-start scenarios involving unseen drugs or targets. To address this issue, we propose LNMGAT, a LapRLS-guided reliable pseudo-negative mining framework coupled with dual graph attention encoders. Instead of relying on experimentally confirmed negative labels or randomly sampled negatives, LNMGAT first applies Laplacian regularized least squares to drug and target similarity graphs to identify low-confidence unlabeled pairs as reliable pseudo-negatives. Drug and target representations are then learned separately on similarity-based k-nearest-neighbor graphs using graph attention networks, and their embeddings are concatenated for MLP-based interaction prediction. Across Yamanishi, Davis, KIBA, and BindingDB benchmarks, LNMGAT achieved the best AUPR in 10 of 16 evaluation settings and ranked within the top two in 14 of 16 settings. In the 12 cold-start settings, LNMGAT obtained the best AUPR in 9 cases, with absolute AUPR gains over the strongest baseline of up to 0.016 in pair cold-start prediction. External evaluation on DrugBank positive interactions and SwissDock-based molecular docking further provided database-level and in silico support for the plausibility of high-ranked predictions. Nevertheless, the biological validation in this study remains computational and database-based; no wet-lab binding assay was performed. LNMGAT therefore provides a competitive and interpretable framework for DTI prediction under negative-label uncertainty, while further experimental validation is required for its top-ranked candidates.

The production of green hydrogen through water splitting requires highly efficient electrocatalysts, but the current trial-and-error-based synthesis or discovery is time-consuming, costly and resource-intensive. Machine learning (ML) provides a powerful, data-driven alternative that can model complex structure-activity relationships across large chemical spaces at orders-of-magnitude speed. This review systematically overviews the life cycle of the electrocatalyst research and development application of ML. First, the thermodynamic and kinetic principles of the hydrogen and oxygen evolution reactions are summarised, along with some well-adopted and accepted activity descriptors. Then we explore data sources, featurization approaches, and algorithms, and discuss the model space, from a simple interpretable model to a graph neural network to a generative model, in the context of the ML toolkit. Strategic applications are discussed for high-throughput virtual screening of alloys and single-atom catalysts, as well as multifunctional activity prediction for overall water splitting, and stability optimisation under operating conditions. The topic of emerging frontiers is highlighted, including high-entropy alloys, amorphous materials, and linking atomic-scale understanding to device-level performance through integration with density functional theory. Finally, the problems of data scarcity, model interpretability and the discrepancy between computational predictions and industrial implementation are discussed, along with future directions for closed-loop discovery and self-driving laboratories. Incorporating ML into electrocatalyst design and combining it with autonomous experimentation will revolutionise this process, from simulation to energy solution, dramatically speeding it up.

We study a class of multi-stage stochastic programs, which incorporate modeling features from Markov decision processes (MDPs). This class includes structured MDPs with continuous action and state spaces. We extend policy graphs to include decision-dependent uncertainty for one-step transition probabilities as well as a limited form of statistical learning. We focus on the expressiveness of our modeling approach, illustrating ideas with a series of examples of increasing complexity. As a solution method, we develop new variants of stochastic dual dynamic programming, including approximations to handle non-convexities.

Spatial transcriptomics enables molecular mapping of ischemic stroke tissue, but spatial domain identification is challenging when injury disrupts normal tissue geometry. Methods relying on a single spatial-proximity graph cannot connect physically distant spots that share damage-associated transcriptional programs.

Autism spectrum disorder (ASD) is a complex neurodevelopmental condition with a broad spectrum of symptoms, which makes timely and accurate diagnosis challenging. The development of machine learning (ML) and deep learning (DL) has created opportunities for automated ASD screening and detection. This systematic review focuses on the analyses of 59 peer-reviewed studies on unimodal and multimodal approaches to ASD detection that were published between 2019 and 2025. The results demonstrated that classical ML algorithms (such as logistic regression [LR], support vector machines [SVM] and random forests [RF]) and DL models (convolutional neural networks [CNN], recurrent neural networks (RNN) and transformers) were used to assess the accuracy of the diagnosis for a variety of data modalities ranging from behavioural measures to neuroimaging, electroencephalography (EEG), eye tracking and speech, with accuracy from 68% to 99%. A careful examination of these studies, however, shows that they share certain common flaws, including small sample size, demographic bias, overfitting and absence of external validation. Hybrid multimodal frameworks have been shown to yield consistent performance improvements over unimodal frameworks, with accuracies of 95%-99% achieved through attention, graph-based learning and hybrid fusion approaches. This review highlights four major points: (1) a critical review of dataset ethics and validity, even for non-clinical facial image datasets; (2) an architectural comparison of multimodal fusion strategies (early fusion, late fusion and hybrid fusion) focusing on computational complexity and clinical applicability; (3) a quantitative summarization of the performance trends by modalities and sample size; and (4) a structured review of indicators of reproducibility and regulatory hurdles for clinical translation. This review suggests the need to develop large, well-balanced datasets, the application of explainable AI (XAI) techniques, standardization (e.g., brain imaging data structure [BIDS]) and regulatory guidelines for facilitating the clinical translation of ASD detection systems.

Eukaryotic genomes self-organize into diverse three-dimensional (3D) chromatin conformations through intranuclear long-range interactions, yet it remains challenging to interpret how concurrent interactions combine along a chromatin polymer to shape measurable conformational readouts. Building on a minimal harmonic polymer model and the Gaussian covariance formalism, we present a 3D genome circuit representation that reorganizes the covariance-derived effective interaction strength (EIS) of a target locus pair into motif-level interaction patterns. In this representation, the graph topology of a chromatin interaction network determines the arrangement of circuit operations that summarize how local loop motifs contribute to EIS through series-like, parallel-like, or more general coupled-network combinations. The resulting EIS is then connected to experimentally accessible statistics, including pairwise contact probabilities and spatial-distance distributions from the Gaussian equilibrium ensemble, and loop stability through a first-passage description. Beyond single-pair readouts, we derive a closed-form Pearson correlation coefficient between two inter-locus spatial distances in a general harmonic network, providing an analytical way to quantify coordinated proximity changes between two locus pairs. We apply the framework to (i) estimate the effective Shh-ZRS interaction strength in a preformed CTCF-dependent loop configuration and (ii) quantify enhancer-promoter proximity coordination in a shared-enhancer configuration. This motif-level circuit representation provides physics-grounded design rules for interpreting how local interaction topology, interaction strength, and perturbations such as anchor deletion or tether addition alter EIS and associated 3D conformational readouts.