Per- and polyfluoroalkyl substances (PFAS) constitute a large and structurally diverse class of man-made chemicals. Their strong carbon-fluorine (C-F) bonds confer high environmental persistence, bioaccumulation, and various associated toxicities. As amphiphilic compounds, most PFAS bind to proteins and accumulate in protein-rich tissues, with such bioaccumulation exerting significant adverse impacts on human health. Accurate evaluation of the binding status between PFAS and proteins constitutes an essential step in health risk assessment. Traditional experiments and certain modeling approaches for analyzing PFAS bioaccumulation suffer from drawbacks such as time-consuming processes, high costs, or inadequate capture of molecular structural information, while existing machine learning-based prediction methods rely on single molecular representation, making it difficult to comprehensively encode the structural information on PFAS. Here, we propose MURNet, a multirepresentation fusion network model integrating chemical descriptors, 2D molecular graphs, and molecular fingerprints to predict PFAS-plasma protein binding. Compared with the state-of-the-art baseline models, MURNet achieves the optimal comprehensive performance. The multirepresentation fusion strategy generates higher-quality molecular features. Tanimoto similarity applicability domain analysis demonstrates MURNet's capability to reliably predict PFAS homologues. Case studies reveal the effectiveness of MURNet in screening PFAS with potential binding affinity to human serum albumin (HSA).

In the early stages of drug discovery, predicting drug-target affinity is a crucial task. Due to the vast scale of genomic and chemical spaces, traditional biological methods are time-consuming, labor-intensive, and resource-demanding. As a result, machine learning-based computational methods have emerged to narrow down the pool of drug candidates. However, machine learning approaches still face several challenges in practical applications, particularly the scarcity of labeled samples and poor model generalization capability. To address these issues, this paper proposes a novel drug-target affinity prediction model, termed MetaBayes-DTA, based on an uncertainty-aware meta-learning framework. The model integrates the few-shot rapid adaptation capability of meta-learning with an uncertainty quantification mechanism to enhance prediction accuracy and reliability. MetaBayes-DTA is evaluated on two benchmark datasets, DAVIS and KIBA. Experimental results demonstrate that the proposed model outperforms existing methods.

Single-cell sequencing has transformed our understanding of cellular heterogeneity, enabling the construction of multi-omics atlases through data integration. However, conventional atlas updates require full reintegration of all datasets, creating scalability challenges that limit the timeliness and adaptability of biomedical research. Here we present multimodal integration with continual learning (MIRACLE), an online learning framework for scalable multimodal integration. Using dynamic architecture adaptation and data rehearsal, MIRACLE continually integrates diverse datasets while preserving biological fidelity. Across evaluations, MIRACLE achieves accurate online integration with substantially improved efficiency, refining and expanding atlases with new cross-modal, cross-tissue and cross-disease data. Applied to respiratory infections, it reveals both shared and pathogen-specific immune mechanisms in coronavirus disease 2019, influenza A and tuberculosis. Overall, MIRACLE provides an efficient and collaborative solution for the continual integration, sharing and exploration of biological knowledge.

MOTIVATION:Predicting the thermodynamic stability of proteins upon single-point mutations is a pivotal step in both protein engineering and medicine. In the study of predicting protein thermodynamic stability, various computational methods, whether they extract features at the local-level or global-level, exhibit their respective advantages and limitations. To leverage the advantages of both features, we developed MuFaDDG, a novel sequence-based method that integrated multiscale feature fusion for improved prediction of protein stability changes (ΔΔG). RESULTS:MuFaDDG achieves comparable performance on the S669 benchmark, demonstrating strong capabilities in stabilizing mutations. Notably, it shows a significant advantage in the ACC metric, with values of 0.75, 0.88, and 0.81 on the direct, reverse, and overall datasets of the CAGI5 Challenge's Frataxin, respectively. Furthermore, our method outperforms leading sequence-based approaches including THPLM, DDGemb, DDGun, and INPS-Seq on protein Myoglobin stability prediction. Additionally, MuFaDDG demonstrates exceptional predictive performance with higher PCC and ACC on the protein ThreeFoil, which is uncurated by FireProtDB and ProThermDB databases. AVAILABILITY AND IMPLEMENTATION:The source code and data are available at https://github.com/PengjiaMa23/MuFaDDG.

The evolution of SARS-CoV-2 has been driven by successive globally circulating waves, including the Alpha and Delta lineages, early Omicron (BA.1-BA.5), XBB, and the recently dominant JN.1 lineages. Although the marked advantage in fitness of early Omicron over Delta lineages has been recognized, there is a lack of systematic evaluation of SARS-CoV-2 fitness across 2020 to 2025. Here, we analyzed 15.23 million SARS-CoV-2 genomes available through May 2025. The accumulation of mutations in the spike protein of the virus has continued to accelerate over time, whereas the trend slowed in the other viral proteins. Using a Bayesian genomic-epidemiological framework, we estimated that lineage fitness increased approximately linearly from 2021 to 2025. Notably, JN.1 lineages exhibited a significantly higher rate of fitness gain than their predecessor XBB and earlier Omicron lineages. We further analyzed characteristic mutations of JN.1 and found that those in the receptor-binding domain were associated with larger alterations in residue hydropathy, charge, and structural surface exposure relative to other lineages. These findings suggest JN.1 as a distinct evolutionary stage and underscore the importance of sustained genomic surveillance.

Background:Periodontitis is a chronic infectious disease caused by bacteria, which leads to destruction of periodontal tissues, tooth loss, and systemic complications. Conventional treatments often fail to counteract the suppression of periodontal tissue regeneration caused by the persistent inflammatory microenvironment. Scope:This review focuses on the application of functionalized nanozymes in the treatment of periodontitis, covering the classification of nanozymes, their mechanisms of action, and recent advances in nanozyme-based therapeutic strategies. Key Findings:Functionalized nanozymes, with their multiple bioactivities including antibacterial, antioxidant, and osteogenic stimulation, represent a promising complement to existing therapeutic approaches. Their sophisticated designs enhance biofilm eradication, modulate immune responses, and facilitate tissue regeneration, thereby overcoming key limitations of existing periodontal treatments. Conclusion:Functionalized nanozymes, particularly motor-based composite nanozymes, show great promise for periodontitis treatment due to their self-propulsion and multifunctional design (antibacterial, antioxidant, osteogenic). However, long-term biosafety, especially metal-ion accumulation, remains the key bottleneck for clinical translation. With continued optimization and standardized safety evaluation, these nanozyme platforms may open a new precision therapy avenue for drug-resistant refractory periodontitis.

Synergistic prediction of anti-cancer drug combinations is a key problem in combination therapy and is important for improving cancer treatment efficacy while reducing toxic side effects. However, existing deep learning methods often rely on simple molecular fingerprints or coarse concatenation-based feature fusion, making it difficult to fully capture complex interactions between drugs and cancer cell lines. To address these limitations, this paper proposes SynerMolNet, an anti-cancer drug combination synergy prediction model based on pre-trained molecular representations and attention-aware feature fusion. The model leverages pre-trained three-dimensional molecular representations to extract deep structural features of drugs. It also reduces the dimensionality of high-dimensional cell line gene expression profiles and uses a shared-weight encoder with attention-aware fusion to capture fine-grained drug-drug and drug-cell line interactions. Multiple groups of experiments demonstrate the effectiveness and generalization capability of the proposed method, indicating its potential value for computational screening of anti-cancer drug combinations.

Designing drug-like molecules that satisfy both properties requirements and structural constraints remains challenging. Current generative approaches typically reduce properties to numerical constraints and treat molecular structures as deterministic functions of properties, failing to capture complex, nonlinear structure-property relationships and limiting generative diversity and controllability. Therefore, we propose BiSP-GP, a bidirectional structure-property generative pretraining framework that unifies molecular generation and property prediction as a single sequence modeling task. BiSP-GP serializes continuous properties into semantic token sequences for joint modeling with molecular structures in a shared sequence space. Molecular generation and property prediction are cast as autoregressive sequence modeling tasks, with a cross-modal decoder supporting bidirectional mapping. The framework also incorporates scaffolds as conditional inputs to guide structure generation. Experiments demonstrate that BiSP-GP achieves strong performance gconditional molecular generation, property prediction, and downstream tasks. A case study on PAK1 further validates the model's generative ability and shows improved binding capacity in molecular docking evaluation.

Accurate and robust prediction of drug-target affinity (DTA) plays a critical role in drug discovery. While deep learning has advanced DTA prediction, existing methods struggle with limited training data and poor generalization. In this study, we propose AdaMBind, a novel DTA prediction model based on meta-learning framework with an adaptive task module designed for low-data scenarios. It employs a dynamic "easy-to-hard" task scheduling mechanism to enhance training efficiency and robustness. Experimental results on three benchmark datasets demonstrate that AdaMBind outperforms 8 baseline models in predicting affinity for unseen targets, particularly under few-shot conditions. Under stringent data constraints, the model successfully identifies high-affinity compounds for ESR and TP53, achieving outstanding virtual screening performance. Furthermore, when applied to inhibitor discovery against FLT3 for acute myeloid leukemia, AdaMBind successfully identified candidate compounds with potent inhibitory activity, as verified by preliminary experimental assays. In summary, AdaMBind provides a robust framework for few-shot DTA prediction.

Microbial enzymes constitute a vast reservoir for biocatalyst discovery, yet the functional catalog of enzymes within the microbial proteome remains largely uncharted. Here, we develop RAMER, a pretrained multimodal artificial intelligence (AI) model for enzymes, integrating protein sequences, structures and catalytic reactions. RAMER substantially improves recall for enzyme function annotation at level 4 Enzyme Commission (EC) classes. Applying RAMER to 3.08 billion proteins from marine, freshwater, terrestrial, and extreme microbiomes, we constructed an atlas of 1.04 billion enzymes assigned to over 5600 EC classes, representing nearly 30 times the number of automatically annotated enzymes in UniProtKB/TrEMBL. For microbiomes from distinct environments, enzymes account for 40% to 51.1% of the proteomes and exhibit broadly similar catalytic function profiles. Beyond EC classes, RAMER enables fine-grained organization of specific enzymes in a manner that reflects functional structural domains and conformations, further facilitating enzyme mining and exploration of natural enzyme diversity. With RAMER, we mined and validated fluorinases from Actinomycetota and PETases from the Mariana Trench microbiome, some of which possessed new structural conformations. Together, the atlas provides a comprehensive resource and a perspective of microbial enzymes, and RAMER offers an AI framework for function-driven biocatalyst discovery.

Background Helicobacter pylori (H. pylori) has a highly plastic genome and can generate substantial within-host diversity during chronic gastric colonization. However, the delineation of its within-host subpopulations, particularly regarding the emergence and spread of antibiotic resistance-conferring mutations, remains poorly understood.Results In this study, we enrolled 25 chronic gastritis patients from southern China, collecting multiple isolates from distinct gastric regions. Among them, 14 patients exhibited heterogeneity in antibiotic susceptibility across isolates (heteroresistant), while the remaining 11 showed consistent profiles (homoresistant). Using ultra-deep short- and long-read sequencing, we showed that co-existing H. pylori subpopulations were prevalent in these patients, particularly within the same anatomical niche. Two patients presented mixed infections involving different strains as subpopulations, while others exhibited microevolution from a common ancestor. We reconstructed the subpopulation structures and found that isolates from heteroresistant patients had greater within-host diversity compared to those from homoresistant patients. Notably, subpopulations in the antrum demonstrated higher diversity than those in the gastric corpus and incisura angularis. Through a custom-developed phasing bioinformatics workflow, we resolved subpopulation-level genomic regions and directly observed extensive homologous recombination among them. Importantly, we traced the distribution of levofloxacin- and clarithromycin-associated resistance mutations across subpopulations, which was mainly mediated by recombination.Conclusions To our knowledge, this study provides the first detailed depiction of H. pylori subpopulation distribution within the human stomach, illustrating how recombination drives within-host diversification and contributes to the spread of antibiotic resistance mutations.

Motivation The goal of molecular representation learning is to automate the extraction of molecular features, a critical task in cheminformatics and drug discovery. While pretraining models using multiple views like SMILES, 2D graphs, and 3D conformations have advanced the field, integrating them effectively to produce superior representations remains a challenge.Results To bridge this gap, we propose a novel multi-view molecular pretraining method termed MMPCS, which explicitly factorizes representations into consistency and specific information. Our approach utilizes the Graph Isomorphism Network and the RoBERTa model to encode 2D molecular topological graphs and SMILES sequences, respectively. Each resulting molecular embedding is decomposed into a shared consistency component and a view-specific remainder. An autoencoder then aligns the consistency information across views. The combined consistency and view-specific representations serve as input for downstream tasks, enabling precise and task-aware predictions. When benchmarked against 16 state-of-the-art molecular pretraining methods, MMPCS achieved the highest average performance across both classification and regression tasks for molecular property prediction. It also delivered outstanding results in predicting drug-target binding affinity and cancer drug response, demonstrating its robustness and broad applicability. Additionally, a case study on the SARS-CoV-2 Omicron variant highlights the potential of MMPCS in facilitating drug repurposing efforts.Availability and implementation The source code and datasets supporting this study are publicly available at GitHub (https://github.com/xmubiocode/MMPCS) and Zenodo (https://doi.org/10.5281/zenodo.18182748).

Deep learning has made significant progress in drug-target interaction (DTI) prediction. However, most existing approaches are developed using fixed datasets, which limits their applicability and generalizability due to the restricted scale of current datasets. This leads to significant discrepancies across the vast chemical space, particularly affecting model performance for previously unseen drugs or targets. In this study, we introduce ASCENT, an active transfer learning framework for DTI prediction. ASCENT utilizes an adaptive active learning strategy to expand datasets by dynamically selecting and annotating the most representative and uncertain samples based on model performance. To improve transferability, an entropy-based adversarial method was incorporated to align feature spaces between source and target domains during training. These innovations enable ASCENT to efficiently capture patterns across vast chemical spaces, thereby enhancing predictive accuracy while reducing annotation costs. Experimental results validate ASCENT's superiority in cross-domain applications, demonstrating its capacity to rapidly and effectively explore chemical diversity. Notably, ASCENT achieves desired performance levels while decreasing annotation expenditures by approximately 20%. Additionally, four representative case studies underscore ASCENT's potential in novel drug discovery and drug repurposing. These results highlight ASCENT as a valuable methodological advancement that supports accelerated drug development and provides new perspectives on DTI prediction.

Nanopore sequencing provides long and ultra-long reads that are valuable for structural variation (SV) detection and genome assembly. However, false-positive chimeric reads can arise and interfere with somatic SV calling. Here, we show that ligation-based library preparation generates false-positive chimeric reads, particularly inverted repeats, in both microbial and human DNA standards. The proportion of inverted repeats ranged from 0.18% to 7.33%, exceeding those observed in rapid and modified rapid preparations. Analysis of raw electrical signals revealed a characteristic smoothed segment at junction sites in nearly half of these chimeric reads. Based on these features, we developed a ResNet-based deep learning classifier to identify false-positive chimeric reads. The model achieved high accuracy in both human and microbial datasets and substantially reduced SV calling errors after filtering. These results demonstrate that library preparation-induced chimeric reads can be effectively detected and mitigated, improving the reliability of nanoporebased SV analysis.

Drug-induced liver injury (DILI) is a major cause of drug development failure and post-marketing withdrawal. Accurate computational prediction of hepatotoxicity is hindered by complex biological mechanisms and scarce labeled toxicity data. Although pretrained molecular language models like ChemBERTa perform well in molecular property prediction, their generalization ability for small-sample DILI prediction remains underexplored. Here, we systematically compared traditional molecular fingerprint-based machine learning methods and ChemBERTa-based models for DILI classification on the DILIst dataset. Canonical SMILES from PubChem were used to generate Morgan fingerprints and ChemBERTa embeddings. We evaluated Random Forest, XGBoost, full fine-tuning, frozen encoder transfer learning, and embedding-based classifiers under both random and scaffold data splits. Results showed that Morgan fingerprints combined with Random Forest achieved the best performance, with ROC-AUC of 0.783 and PR-AUC of 0.849 under random split. Scaffold split markedly degraded the performance of all models, indicating poor generalization to unseen chemical scaffolds. ChemBERTa embedding-based classifiers outperformed end-to-end fine-tuning, suggesting that pretrained representations are better used as fixed feature extractors under limited labeled DILI data. Further SHAP analysis detected key toxicity-related molecular fragments, and t-SNE showed insufficient latent-space separation between DILI-positive and negative compounds. Our results confirm that traditional fingerprint-based machine learning remains highly competitive for small-sample hepatotoxicity prediction, and this work provides a reliable computational framework for early drug safety assessment.

Nanopore adaptive sequencing enables real-time target enrichment, yet current deep-learning methods require costly, sample-specific experimental training data. To address this, we developed GANBase, a genome-guided generative adversarial learning framework, which is trained exclusively on reference sequences and incorporates a Monte Carlo Tree Search-based Rollout strategy for model training. GANBase demonstrates robust performance in target enrichment and host depletion across diverse scenarios. In live adaptive sequencing experiments, it remains effective despite significant pore loss or flow cell version updates, providing a data-independent solution that significantly expands the utility of real-time targeted sequencing.

Machine learning has been successfully applied to drug combination prediction in recent years. However, in some situations, the class imbalance problem still shows highly negative impacts on the modeling process, which cannot be directly handled by traditional methods. In addition, the interpretability of models is another key point for biological and medical experts. In this study, a clustering-based oblique decision tree (MOCT) algorithm is proposed to extract interpretable knowledge for the multi-class datasets. It firstly clusters samples of different classes, and then a proper feature subspace is generated to split data and forms a nonleaf node. Unlike traditional decision trees, our MOCT only grows one none-leaf node in each layer to generate a concise tree structure. Datasets of drug combinations were collected from three cell lines with three classes (Additive, Antagonism, and Synergy) in experiments, and the results show that our MOCT algorithm is superior to other methods with better interpretability.

Prediction of protein conformational ensembles is essential for understanding protein functions and advancing drug design. Molecular dynamics (MD) simulation is computationally intensive for prediction; generative artificial intelligence (AI) frameworks offer a promising alternative. To ensure model performance, existing methods continuously expand training data, including experimental structures, AlphaFold predictions, and MD trajectories; however, model performance still needs improvement. Here, we demonstrated that incorporating protein complex interactions enhances model performance more effectively than relying solely on single-chain structures. Supplementing single-chain data with a mere 10% of complex data improved the performance of monomer-trained models by 6.64% in PwD-JSD and 22.87% in Rg-JSD. Building upon this insight, we introduce DynoM, a generative framework trained on 5,502 in-house all-atom MD trajectories of protein complexes. DynoM demonstrates high robustness in generating monomeric ensembles; notably, for multi-domain proteins, the model effectively suppresses structural hallucinations in inter-domain positioning by capturing the long-range physical constraints inherent in complex interactions. Overall, our findings underscore the pivotal role of complex interactions in predicting accurate conformational ensembles and provide a robust tool for protein modeling research.