Non-coding RNAs (ncRNAs), particularly long non-coding RNAs (lncRNAs) and microRNAs (miRNAs), are important regulators of gene expression and are closely involved in disease pathogenesis. Therefore, identifying ncRNA-disease associations is essential for clarifying disease mechanisms. Because lncRNAs and miRNAs frequently regulate each other in cells, predicting miRNA-disease associations (MDA) and lncRNA-disease associations (LDA) is inherently interconnected. However, many existing computational approaches still treat these tasks as independent problems, which limits their ability to capture cross-task biological signals. In addition, most models use fixed hyperparameters, which may be suboptimal as training dynamics and data characteristics change. To address these issues, we propose RL-DMGLMD (Reinforcement Learning-enhanced Dual-view Multi-task Graph learning for LncRNA-MiRNA-Disease association prediction). RL-DMGLMD contains three key components: (1) a Soft Actor-Critic (SAC) controller that adaptively tunes hyperparameters during training by monitoring loss, validation performance, and gradient information; (2) a unified multi-task framework that jointly predicts LDA, MDA, and lncRNA-miRNA interactions (LMI) using shared encoders with task-specific decoders to enable knowledge transfer; and (3) a dual-view multi-head Graph Attention Network (GAT) that learns from both heterogeneous interaction graphs and attribute graphs to capture relation-specific importance. RL-DMGLMD achieves AUROC values of 0.9900 / 0.9872 / 0.9867 on Dataset 1 and 0.9946 / 0.9903 / 0.9954 on Dataset 2 for LDA, MDA, and LMI, respectively. These results outperform state-of-the-art baselines and support RL-DMGLMD as a practical tool for biomarker discovery and therapeutic target prioritization.

The rapid expansion of enzyme reaction literature has created a major bottleneck in database curation, leaving vast amounts of enzyme-substrate-condition relationships unstructured and inaccessible for DL-driven modeling. How to fully utilize the enzymatic reaction data has been an important task for future accurate enzyme activity prediction models. Current deep learning (DL)-based data extraction models heavily rely on large language models (LLMs) without a fidelity check and the ability to continuously evolve. To address these issues, we developed zERExtractor (Zelixir's Enzyme Reaction Data Extractor), an accuracy-oriented and extensible platform for extracting enzyme-catalyzed reaction data from scientific publications. This system offers a unified multimodal information extraction framework (covering molecular reaction diagrams, tables, and texts) to integrate enzymatic reaction descriptors into structured storage. We employ fine-tuned large LLMs together with DL in a human-in-the-loop pipeline that evolves through data fidelity validation by experts and active learning. Also, zERExtractor achieves 89.9% accuracy in table recognition and over 98% accuracy in molecular image recognition on synthetic data sets, outperforming the strongest baseline by more than 2% and consistently maintaining above 95% on realistic benchmarks. zERExtractor bridges the data gap in enzyme reaction data with a scalable framework for accurate multimodal extraction, advancing DL-driven enzyme modeling and enabling future applications in computational enzymology and biotechnology. The platform is publicly accessible online at https://zpaper.zelixir.com/.

Molecular property prediction is a cornerstone for accelerating drug discovery, providing a computational way to alleviate the low success rate, high cost, and long development cycle of conventional pharmaceutical research and development. However, existing computational methods have obvious limitations: most current deep learning approaches rely on a single molecular view, failing to fully capture the multi-dimensional features of molecular structures; they mainly focus on intra-molecular features while ignoring inter molecular information and cross-task correlations; and limited labeled data severely impairs their generalization to novel molecules. To address these issues, we propose MSMPP, a multi-scale, multi-view fusion framework for molecular property prediction that learns intra- and inter scale features simultaneously. For intra-scale feature learning, MSMPP integrates TxGemma-enhanced 1D sequence representations, Graph Transformer-derived 2D topological graph features, and Uni-Mol-derived 3D molecular conformational features. The Graph Transformer models long-range atomic dependencies, while the two pretrained models provide task-agnostic molecular prior knowledge from large-scale pretraining corpora, thereby improving generalization to novel molecules. For inter-scale feature learning, MSMPP constructs an inter-molecular graph (IMG) that explicitly models global pairwise interactions among chemically similar molecules and also extracts the cross-task features. Evaluations on eight MoleculeNet datasets show that MSMPP significantly outperforms state-of-the-art models, demonstrating its effectiveness in integrating multi-view intra-molecular features, inter-molecular features and cross-task information. Overall, MSMPP provides a competitive tool for molecular property prediction and supports the acceleration of drug discovery workflows.

Multi-type epileptiform events, including spike and ripple activities in electroencephalogram (EEG), are important biomarkers for seizure detection; however, their accurate detection is challenging due to complex spatio-temporal dynamics and multi-scale oscillatory patterns. To address these challenges, we propose a unified framework that integrates clinically motivated multi-band decomposition with dynamic spatio-temporal representation learning for accurate classification of EEG signals. Specifically, EEG signals are first decomposed into multiple frequency branches, followed by a branch-aware wavelet enhancement module to emphasize transient oscillatory patterns. A multi-scale temporal convolution module is then employed to capture diverse temporal dynamics across different time scales. Finally, a dynamic spatio-temporal graph reasoning module constructs sample-adaptive spatial and temporal graphs and performs joint propagation to model inter-channel interactions and temporal dependencies. Extensive experiments conducted on both a private clinical dataset and the public TUEV dataset demonstrate that the proposed method consistently outperforms existing state-of-the-art approaches in epileptic event detection. These results highlight its effectiveness in capturing clinically relevant EEG patterns and its potential for assisting automated epilepsy diagnosis.

Missense mutations are common in the coding genome and can alter protein functions. Distinguishing pathogenic from benign variants remains challenging despite computational advances. In the present work, we introduce EMMVEP, an ensemble-based approach designed for predicting the effects of protein missense mutations. EMMVEP leverages categorical boosting to integrate different types of features: one-hot encoding from protein sequence, physicochemical and environment properties extracted from AlphaFold database, and allele frequency information from gnomAD. When evaluated on a benchmark dataset with 112,832 clinical significance labels, our method achieved AUC and AUPR of 0.907 and 0.879, outperforming 20 general VEP methods. To aid in the identification of pathogenic mutations among the vast number of rare variants discovered through large-scale sequencing studies, we provide the pathogenicity probabilities of 216 million potential amino acid substitutions in 19,233 human protein-encoding genes. Our work demonstrates that EMMVEP can offer valuable independent insights for missense mutation interpretation in proteins, with significant applicability in both research and clinical contexts.

Autism spectrum disorder (ASD) encompasses a range of neurodevelopmental conditions characterized by impairments in social interaction, communication, and behavior. Early detection is critical for timely intervention and improved outcomes; however, conventional approaches remain subjective, labor-intensive, and dependent on highly trained professionals. Facial image-based ASD research also faces challenges such as inter-class resemblance, intra-class variation, and pose differences. To address these issues, we propose the Multi-Aspect Cross-Fusion Transformer (MACFUT), a network that integrates facial landmark and image features through a multi-level cross-fusion mechanism for early ASD screening. The architecture incorporates spatial attention to highlight salient facial regions and a global-local cross-fusion transformer encoder to capture detailed feature relationships. We further analyze facial landmark distances, vectors, and angles to reveal developmental differences and enhance clinical interpretability. The proposed model is trained on the ASD Facial Image Dataset (AFID) using 10-fold cross-validation, achieving 89.56% accuracy and an area under the receiver operating characteristic curve (AUC) of 95.36%. To evaluate cross-dataset generalization, the trained model is tested on an independent dataset from Bangladesh (BACFED), yielding 86.51% accuracy and an AUC of 94.78%. The model also highlights diagnostic-group and sex-based differences in facial features among children. MACFUT provides an interpretable and accessible early ASD screening tool that demonstrates robustness across heterogeneous datasets. This method outperforms existing approaches and may help identify children who could benefit from further clinical evaluation in resource-limited and telehealth settings.

The significantly negative impact of marine invasive species underscores the need to understand the dynamics of invasion success. MicroRNAs (miRNAs) play a crucial role in regulating gene expression in response to stresses during invasions. Using the invasive tunicate Ciona robusta as a model, here we aim to study intragenic miRNA–host gene co-expression and functional regulation in response to recurrent salinity challenges. Despite genomic nestedness, only 9

Biochemical reaction diagrams encode molecular entities, transformations, and experimental conditions in tightly coupled visual-symbolic structures, requiring reasoning beyond element-level recognition. We formulate biochemical reaction diagram question answering as a hierarchical multimodal reasoning task and introduce ReactionQA, a benchmark spanning element identification to consistency verification. We further propose BioRxnReasoner, a multi-stage multi-agent framework that integrates structured reaction representation, question decomposition, multimodal perception, and answer verification. BioRxnReasoner consistently outperforms parsing-based pipelines and end-to-end multimodal models, achieving 83.2 https://github.com/cs-ai4sci/BioRxnReasoner .

Nitrous oxide (N2O) is a potent greenhouse gas, with agricultural activities representing its major source. However, the emission mechanism of nitrous oxide efficient by agricultural activities has not yet been fully studied. This study employs metagenomic analysis to elucidate the microbial community structure and functional potential associated with N2O emissions in river and ditch systems of the Yangtze River Delta. The N2O dissolved concentration in the rivers (0.08 +/- 0.03 mu mol NL-1) was significantly lower than that in the ditches (0.21 +/- 0.14 mu mol NL-1) (P < 0.05). According to eight wind-based models, agricultural ditches emissions were 3.53-4.70 times higher than those of the rivers. All models significantly overestimated fluxes (P < 0.05), revealing a systematic overestimation of EF values when using the Intergovernmental Panel on Climate Change (IPCC) methodology. Particulate organic carbon supported microbial activity by providing energy and adhesion sites, while electrical conductivity (EC) served as an indicator of ion inputs from surrounding land use, serving as a critical abiotic driver of EF values in the ditches. The co-occurrence network showed that denitrification genes (norB, nirS, nosZ) formed a tightly clustered subnetwork exclusively in the ditches, indicating broader nitrification niches and stronger functional coupling among denitrifiers in these systems. Metagenomic evidence revealed that EF value correlated significantly with denitrification genes, notably napAB, nirK, norBC and nirK/nosZ (P < 0.05), underscoring denitrification as the primary biotic driver of N2O production. These findings demonstrate the value of metagenomic approaches in revealing microbial mechanisms behind N2O emissions and support the development of more accurate, EF estimates for greenhouse gas inventories in agricultural landscapes.

The identification of latent DNA binding domains presents both significant scientific value and analytical complexity, given the extensive diversity within biological compound datasets. To address this challenge, our research introduces weighted deep forest (WeighDF), a novel computational framework integrating hybrid feature representation with adaptive multi-granularity scanning analysis. This approach dynamically weights features across scanning windows using learnable attenuation coefficients, which amplifies key sequence patterns and suppresses background noise. For comprehensive prediction of diverse DNA binding patterns, we further develop decision learning predictive algorithm for binding sites (DecLPABS), an ensemble architecture combining WeighDF's adaptive scanning with meta-learner integration strategies. This dual-phase system demonstrates superior versatility in handling both categorical classification and continuous regression problems. Empirical validation across heterogeneous datasets reveals DecLPABS's enhanced predictive capability, achieving 0.8979 accuracy through optimized feature-space partitioning.

Allostery mediates long-range regulation in proteins and underlies many biologically and pharmacologically important processes. However, predicting the pathogenicity of mutations at allosteric sites remains challenging because their impacts are often indirect and mechanism-dependent, and labeled data for this setting are scarce. We present AlloMut, a dual-view self-supervised framework that learns mutation-pair representations from unlabeled allosteric-site mutations using a consistency objective and a ranking objective defined by mutation-induced energy differences, and leverages the learned encoder for downstream pathogenicity prediction. In a non-overlapping protein-wise setup, AlloMut was pretrained on 166 proteins with 5,941 annotated allosteric sites and evaluated on 461 labeled mutations across 58 proteins. On the independent test set, AlloMut achieved an AUC of 0.833, an AUPR of 0.919 and an MCC of 0.548, showing competitive performance against 11 existing methods. These results highlight the value of combining self-supervised pretraining with structure-derived priors for label-scarce allosteric variant prioritization.

Protein missense mutations, which involve single amino acid substitutions, can remarkably alter protein structure and function, thereby potentially leading to a broad spectrum of diseases. Accurate prediction of the pathogenicity of such mutations is crucial for elucidating disease mechanisms and promoting the advancement of precision medicine. While computational approaches have emerged as efficient alternatives to costly experimental methodologies, existing techniques neglect the structural information of variant-type proteins and thus often fail to effectively integrate multi-modal protein features. Here, we propose HiFuseVEP, a hierarchical multi-modal fusion framework designed to integrate protein sequence and structural features extracted from protein language models. This framework comprises three core components: a multi-modal feature encoding module based on weight-standardized convolutions and Transformer encoder, a cross-modal adaptive feature fusion module, and a dynamic classification prediction module equipped with channel attention in residual blocks. We constructed a comprehensive dataset containing 89,834 missense mutations across 5,342 human proteins from the ClinVar database. Extensive experimental results demonstrate that HiFuseVEP achieves state-of-the-art performance with an AUC of 0.913 and an AUPR of 0.872, significantly outperforming 17 existing methods. Ablation studies further validate the effectiveness of multi-modal feature fusion and the contribution of each architectural component. Our findings highlight that accurate pathogenicity prediction relies on the simultaneous integration of local mutation-site features and global protein-level representations derived from both wild-type and variant-type sequences and structures. The data and code for HiFuseVEP are available at https://github.com/mjcoo/HiFuseVEP for academic use.

This study surveyed the Xiaoheishan Island artificial reef area in China’s Bohai Sea (deployed in 2014) through year-round ecological surveys conducted in 2023. Using environmental DNA metabarcoding technology, we systematically studied the ecological restoration effects and underlying mechanisms of the artificial reef area. The results revealed 91 genetically identified fish species in the artificial reef area, compared to 87 species in the control area, indicating that artificial reefs play a positive role in restoring biodiversity. Principal Component Analysis demonstrated significant seasonal and spatial differences in biological distribution between the artificial reef and control areas (P<0.05). Specifically, the artificial reef area is dominated by small demersal rock-associated fishes such as Gobiidae, Scorpaeniformes, Lateolabrax japonicus, and Hexagrammos otakii; in contrast, the control area features predominantly pelagic and soft-bottom species including Liza haematocheila, Engraulis japonicus, Mugil cephalus, Paralichthys olivaceus, and Scomberomorus niphonius. By Canonical Correspondence Analysis, we propose that seasonal migration driven by water temperature and species’ life-history strategies promote habitat partitioning between the artificial reef and control areas, which in turn influences the distribution of dissolved oxygen and NH 4 + in the water. Through the correlation analyses, it is found that although chlorophyll levels were higher in control area, the category and diversity indices of plankton showed marked declines compared to the artificial reef area. We believe that, compared to control areas, artificial reefs can increase the biodiversity of smaller organisms but cannot support large predators and higher biomass.

The continuous development of next-generation sequencing (NGS) technology has led to extensive and frequent use of genomic analysis in cancer research. The associated production of large-scale NGS datasets establishes the need for high-precision somatic variant calling methods that are highly optimized on commonly used hardware platforms. We present RabbitVar ( https://github.com/LeiHaoa/RabbitVar ), a scalable variant caller that can detect small somatic variants from paired tumor/normal NGS data on modern multi-core CPUs. Our approach combines candidate-finding and machine-learning-based filtering strategies with optimized data structures and multi-threading to achieve both high accuracy and efficiency. We have compared the performance of RabbitVar to leading state-of-the-art callers (Strelka2, Mutect2, NeuSomatic, VarDict, VarScan2) on real-world HCC1395 breast cancer datasets under different sequencing conditions and contamination rates. The evaluation results demonstrate that RabbitVar achieves highly competitive F1-scores when calling SNVs. Moreover, when calling the more challenging indel variants, it consistently achieves the highest F1-scores. RabbitVar is able to process a paired tumor and normal whole human genome sequencing datasets with 80x depth in less than 20 minutes on a 48-core workstation outperforming all other tested variant callers in terms of efficiency.

Understanding the mechanisms underlying invasion success is crucial for addressing the rapidly increasing frequency of biological invasions and their escalating ecological and economic impacts worldwide. Phenotypic plasticity plays a crucial role in facilitating invasion success by enabling organisms to respond rapidly to environmental fluctuations. Among molecular regulators of such plasticity, microRNAs (miRNAs) mediate stress adaptation through post-transcriptional gene regulation. Increasing evidence suggests that a single miRNA locus can generate multiple variants (isomiRs), which may diversify regulatory functions and enhance environmental resilience. However, their expression dynamics and functional significance under environmental stress remain largely unexplored during biological invasions. Here, using the highly invasive ascidian Ciona robusta as a model, we examined the dynamic expression and functional divergence of miRNA isoforms in response to osmotic stress. Integrative analysis of miRNAome and transcriptome revealed 10 miRNAs that produced 5′ isomiRs with time-dependent and stress-specific expression patterns. Both canonical miRNAs and their isomiRs targeted overlapping yet distinct gene sets according to in silico predictions, particularly in pathways related to free amino acid metabolism and ion transport. Functional analysis demonstrated that isomiRs underwent neo-functionalization, sub-functionalization, or mixed functional shifts relative to their canonical counterparts, and in some cases exerted opposite regulatory effects on the same target genes. These results indicate that osmotic stress induces rapid diversification and functional transformation of miRNA isoforms, forming a flexible and dynamic regulatory network. Such plasticity in isomiR regulation likely contributes to enhanced stress tolerance and environmental adaptability, thereby promoting invasion success across diverse, harsh, or rapidly changing environments.

The unprecedented generation of biological data and the computational intensity of modern biomedical research demand transformative solutions. High-Performance Computing (HPC) and Artificial Intelligence (AI) have emerged as pivotal technologies driving innovation in bioinformatics. The combination of AI and HPC has revolutionized genomics, drug discovery, and precision medicine. This special issue invited review and research articles on the latest trends in HPC and AI technologies for biomedical and biological research. Overall 8 papers were accepted in this issue covering a wide range of research areas, including virtual drug screening, biological sequence analysis, drug-target interaction/drug-disease association, metagenomics etc.

The morphology of microalgal cells is intrinsically linked to their biochemical functions, enabling the prediction of cellular functions through the analysis of morphological and visual features extracted from microscopy images. In this work, we introduce FusionODE, a multimodal continuous-time framework that predicts biochemical indicators of the mixotrophic chrysophyte Poterioochromonas malhamensis (P. malhamensis) from microscopy image time series with irregular sampling intervals. Given over 29,000 microscopy images, FusionODE segments individual cells, extracts numerical morphological profiles, and fuses them with image representations through heterogeneous cross-modal attention (HCMA). Interval-conditioned Ordinary Differential Equation-Recurrent Neural Network dynamics (ICOD) captures irregular temporal evolution by incorporating multi-scale sinusoidal interval encodings into non-autonomous vector fields, enabling adaptation to observation gaps spanning multiple time scales. Across three biochemical targets over nine time points, FusionODE achieves an average R^2 of 0.952, a 6.85 https://github.com/ShengqiZhou777/FusionODE .

Molecular generation plays a vital role in advancing drug discovery, materials science, and chemical exploration. In this study, we integrated the conditional variational autoencoder (CVAE) with the Wasserstein generative adversarial network (WGAN) and effectively applied this hybrid architecture to molecular generation tasks, introducing a molecular generation framework with conditional generation capabilities known as CCVAN. The model first encodes the data into a latent vector using CVAE, then shares a decoder with the generator of the WGAN to reconstruct molecular SMILES, and trains the discriminator with both the generated and real molecular SMILES to achieve the goal of conditionally generating molecules. Compared to existing methods, CCVAN can generate molecules with specific properties as needed and performs well in terms of validity and novelty. In a case study, we utilized CCVAN for ligand-based and structure-based drug design, enabling the generation of high-binding-affinity molecules for target binding. Overall, the flexibility and effectiveness of CCVAN make it a valuable tool for accelerating the compound discovery process. The source code of CCVAN is publicly available at https://github.com/mjcoo/CCVAN .