Large-scale scientific simulations require significant resources to generate high-resolution time-varying data (TVD). While super-resolution is an efficient post-processing strategy to reduce costs, existing methods rely on a large amount of HR training data, limiting their applicability to diverse simulation scenarios. To address this constraint, we proposed CD-TVD, a novel framework that combines contrastive learning and an improved diffusion-based super-resolution model to achieve accurate 3D super-resolution from limited time-step high-resolution data. During pre-training on historical simulation data, the contrastive encoder and diffusion superresolution modules learn degradation patterns and detailed features of high-resolution and low-resolution samples. In the training phase, the improved diffusion model with a local attention mechanism is fine-tuned using only one newly generated high-resolution timestep, leveraging the degradation knowledge learned by the encoder. This design minimizes the reliance on large-scale high-resolution datasets while maintaining the capability to recover fine-grained details. Experimental results on fluid and atmospheric simulation datasets confirm that CD-TVD delivers accurate and resource-efficient 3D super-resolution, marking a significant advancement in data augmentation for large-scale scientific simulations. The code is available at https://github.com/Xin-Gao-private/CD-TVD.

Computational protein design is often constrained by slow, complex, inaccessible, and highly sophisticated and expert-dependent workflows that hinder its transferrability and generalization power for broader applications. We present ProteinMCP, an agentic AI framework designed to accelerate and democratize protein engineering. ProteinMCP automates end-to-end scientific tasks, delivering dramatic gains in efficiency; for instance, a comprehensive protein fitness modeling workflow was completed in just 11 min. This performance is achieved by an AI agent that intelligently orchestrates a unified ecosystem of 38 specialized tools, made accessible through a model-context-protocol (MCP). A cornerstone of the framework is an automated pipeline that converts existing software into MCP-compliant servers, ensuring the platform is both powerful and perpetually extensible. We further demonstrate its capabilities through the successful autonomous design and selection of high-affinity de novo binders and therapeutic nanobodies. By removing technical barriers, ProteinMCP has the potential to shorten the design-build-test cycle and make advanced computational protein design accessible to the broader scientific community.

A new AI framework autonomously explores single-cell RNA sequencing datasets and produces creative, biologically relevant findings.

Large Reasoning Models (LRMs) generate intermediate reasoning traces that may contain unsafe content, even when their final responses appear safe. Guardrail models are designed to detect and block unsafe content, yet existing benchmarks for unsafe content detection focus primarily on prompts and final responses, leaving reasoning traces largely unexamined. Moreover, these benchmarks typically provide only binary safety labels, without evidence annotations that justify the judgments. To address these limitations, we introduce TRACE, an evidence-grounded safety evaluation benchmark that covers the entire LRM inference pipeline: prompts, reasoning traces, and final responses. TRACE includes prompts in two languages spanning nine risk categories and ten attack strategies. For each prompt, four LRMs generate reasoning traces and final responses, and we annotate the safety of each component and extract supporting evidence from the corresponding source text. Evaluating 18 guardrail models on TRACE reveals that safety judgment for reasoning traces is substantially more challenging than for prompts or final responses, and that current models struggle to accurately extract supporting evidence. These findings highlight the need for guardrail models that can reliably detect and precisely localize unsafe content across the LRM inference pipeline.

Automated cytogenomic analysis has long been limited by narrow task scope, high annotation demands, and poor robustness to real-world complexity. Here, we introduce CHROMA, the first single-chromosome foundation model for cytogenomics that enables comprehensive, cell-level detection of a wide spectrum of chromosomal abnormalities—including both common and ultra-rare types—in a single, unified framework. Pre-trained on over 4 million chromosomal images from more than 84,000 specimens using self-supervised learning, CHROMA achieves robust and comprehensive detection of numerical and structural abnormalities across diverse classes, dramatically reducing expert annotation workload by 40% through efficient label utilization. The model maintains state-of-the-art accuracy even under highly imbalanced data and challenging imaging conditions, supporting reliable deployment as a risk-aware screening and triage tool, particularly in settings with limited expert availability. An integrated risk-control strategy further ensures safe application by automatically flagging uncertain or rare cases for expert review. By bridging foundational AI advances with real-world clinical needs, CHROMA paves the way for scalable, accessible, and precise cytogenomic analysis in both advanced and underserved healthcare environments.

Scientific reasoning remains challenging for open-source models, largely due to the lack of high-quality scientific reasoning data. Existing datasets are often dominated by factual recall or formulaic problem solving, with limited emphasis on mechanism understanding, evidence-grounded reasoning, and hypothesis evaluation. To address this, we introduce SPARK (Scientific Paper Abstracted Reasoning sKeleton), a paper-oriented synthesis framework built on Sci-Base, a large-scale corpus of research papers spanning 10 scientific disciplines. Instead of directly converting papers into question-answer pairs, SPARK treats the claim-evidence-derivation structure of a paper as the fundamental unit of reasoning synthesis. Specifically, SPARK (1) distills each paper into a compact reasoning skeleton capturing its central claims and supporting evidence, enabling self-contained question generation, and (2) synthesizes reasoning tasks from four scientific perspectives: mechanistic reasoning, hypothesis falsification, quantitative derivation, and boundary calibration. A final consistency verification stage further removes unsupported or contradictory outputs. Using this framework, we construct Spark-234K, a scientific reasoning dataset with substantially higher difficulty and diversity than existing resources. Experiments show that Spark-234K consistently outperforms existing scientific reasoning datasets while achieving stronger performance with significantly fewer training samples.

Generative AI (GenAI) has advanced computational pathology through various image translation models. These models synthesize histopathological images from existing ones, facilitating tasks such as color normalization and virtual staining. Current models, while effective, are mostly dedicated to specific source-target domain pairs and lack scalability for multi-domain translations. Here, we introduce His-MMDM, a diffusion model-based framework enabling multi-domain and multi-omics histopathological image translation. His-MMDM is not only effective in performing existing tasks such as transforming cryosectioned images to FFPE ones and virtual immunohistochemical (IHC) staining but can also facilitate knowledge transfer between different tumor types and between primary and metastatic tumors. Additionally, it performs genomics- and/or transcriptomics-guided editing of histopathological images, illustrating the impact of driver mutations and oncogenic pathway alterations on tissue histopathology and educating pathologists to recognize them. These versatile capabilities position His-MMDM as a versatile tool in the GenAI toolkit for future pathologists.

Motivation Structure-based drug design (SBDD) aims to generate ligand molecules that tightly bind to specific protein targets, a critical step in drug discovery. Diffusion models have shown promise for this task, yet existing methods struggle to effectively incorporate protein-ligand interaction priors during generation. Most approaches rely on protein-specific structural priors that remain fixed throughout generation, limiting molecular diversity and failing to capture the dynamic interplay between protein pockets and ligand atoms, which is essential for achieving high binding affinity.Results We propose DPDiff, a disentangled prior-conditioned diffusion model for protein-specific 3D molecular generation. DPDiff introduces two complementary interaction prior networks that capture geometry-based spatial interactions and sequence-based interactions robust to structural noise. During generation, the model dynamically extracts interaction priors using intermediate diffusion predictions and adaptively fuses them via a time-dependent adapter. A disentangled denoising network balances prior guidance with generative flexibility. Experiments on the CrossDocked2020 dataset demonstrate that DPDiff generates molecules with more realistic 3D structures and state-of-the-art binding affinities, achieving an average Vina Dock score of -8.58 and a high affinity ratio of 69.4%, outperforming existing methods while maintaining favorable drug-likeness and synthetic accessibility.Availability and implementation The source code of DPDiff is available at https://github.com/ZerinHwang03/DPDiff.

Acute respiratory distress syndrome (ARDS) remains a major challenge in critical care, with mortality exceeding 40%. Its diagnosis and management depend on multi-step procedures, invasive arterial blood gas analysis, and subjective CT interpretation, often leading to inconsistency, delayed intervention, and increased procedural burden. To address these limitations, we develop AutoARDS, an all-in-one foundation model that transforms routine chest CT into a quantitative platform, enabling integrated and reproducible assessment of diagnosis, progression, oxygenation, physiology, and prognosis within a single, non-invasive workflow, thereby supporting faster and more standardized critical-care decisions. Technically, AutoARDS proposes to employ a multi-task pretraining strategy with adversarial perturbation, distilling routine but unstructured clinical data into unified representations for fine-grained pathological learning. Trained on over 50,000 CT volumes and validated across six medical centers (6,153 individuals), AutoARDS (1) established a reproducible CT-derived biomarker linking morphological injury with disease severity, enabling standardized tracking of pulmonary progression; (2) accurately diagnosed acute respiratory failure and ARDS (AUCs = 0.97 and 0.87), facilitating early recognition and reducing diagnostic delay; (3) directly estimated the P/F ratio (PCC = 0.83), outperforming SpO2-based monitoring for noninvasive severity stratification and ventilation management; and (4) predicted 28-day outcomes (time-averaged AUC = 0.79), providing complementary risk assessment for clinical planning. Further analyses confirm generalizability to ARDS-associated right ventricular dysfunction (AUC = 0.76) and revealed a positive shift image-derived age residuals, reflecting disease-related imaging patterns that resemble pulmonary aging. By bridging visual information with quantitative physiology, AutoARDS exemplifies a scalable blueprint for transforming chest CT into an integrated, quantitative platform for precise and reproducible critical-care management.

Biomolecular interactions underlie core cellular processes and modern drug discovery, yet reliable prediction across interaction types remains limited by the lack of universal interface representations that transfer across molecular domains. Here, we proposed the Common Interface Network (ComIN), a framework that learns interface representations via contrastive learning on interface atom graphs. By training jointly on protein–protein, protein–peptide, and protein–small molecular interactions, ComIN synthesizes a unified embedding space that demonstrates superior performance over existing domain-specific models. Extensive validation across five distinct interaction-centric tasks in drug discovery and immune recognition underscores the broad transferability and robustness of ComIN as a representational tool. Leveraging this universality, we built ComINdex, a cross-domain search engine indexing million-scale interfaces with ComIN-generated representations to enable efficient retrieval in support of scalable biomolecular function analysis and design.

X-ray tomography is widely used across scientific and clinical domains, yet image degradation remains a major obstacle to reliable analysis, particularly under low-dose or data-scarce conditions. Existing restoration methods are typically designed for specific modalities and predefined degradation, limiting their generalizability. Here we show that image restoration can instead be formulated as learning realistic, nonparametric acquisition degradation processes directly from data. We introduce HorusEye, a self-supervised foundation model for X-ray tomography restoration that leverages interslice contrastive pretraining to jointly learn structural priors and degradation without paired supervision or predefined assumptions. Trained on over 100 million images, HorusEye generalizes across diverse modalities, restoration tasks and previously unseen imaging modalities, consistently outperforming task-specific approaches. Extensive evaluations demonstrate improved photon efficiency and recovery of high-frequency information. Clinical studies further demonstrate enhanced detectability of low-contrast anatomy and lesions, as well as improved performance on downstream tasks, highlighting HorusEye as a general postprocessing tool for X-ray tomography.

Genetic perturbation response prediction plays a critical role in virtual cell research, yet the performance of current deep learning models still leaves room for improvement. In this study, we present a prior-guided response inference model (PRIM) that leverages a valuable priori knowledge of gene expression in control cells to model the effects of perturbations at the cellular level. This allows PRIM to predict the amount of change in gene expression after perturbation to better approximate the real situation. Compared with existing deep learning approaches, PRIM achieves superior performance across multiple data sets and notable advantages in predicting combinatorial perturbation responses. Moreover, it is more lightweight than current deep learning models and enables faster forward inference. Importantly, PRIM effectively captures nonadditive genetic interactions and shows the potential to uncover associations between combinatorial perturbations and new biologically meaningful phenotypes. These findings provide new insights into the application of deep learning for predicting cellular responses to genetic perturbations.

With the growing global threat of coronary artery disease (CAD), automated CAD diagnosis techniques based on coronary CT angiography (CCTA) have been developed. However, their clinical applicability remains limited due to the heterogeneity of stenosis and plaque attributes, as well as confounders within the causal relationships of CAD diagnosis. This work introduces the Attribute-Decoupled Intervention Network (ADI-Net), a confounder-free CAD diagnosis framework designed for fine-grained analysis at both the artery and patient levels, aligning with real-world clinical practice. ADI-Net employs an attribute-decoupled representation that effectively captures the heterogeneous features of stenosis and plaque with differential constraints, enabling precise, fine-grained classification. Additionally, the dynamic-updating causal intervention continuously refines confounder banks and applies the Do-expression within a complete causality, ensuring comprehensive, cross-level assessments. Experiments on CCTA datasets from three clinical centers demonstrate that ADI-Net outperforms state-of-the-art methods in cross-level, fine-grained CAD diagnosis, exhibiting superior robustness, domain adaptability, and data efficiency.

In the filamentous fungus Trichoderma reesei, cellulase production is strongly inhibited by carbon catabolite repression (CCR) mediated by Cre1 and Ace1, and phosphorylation is a crucial post-translational modification that regulates the activity of these two repressors. With the aim of finding approaches to reduce CCR, we used phosphoproteomics to identify two evolutionarily conserved phosphorylation sites, S204 and S145, in Cre1 and Ace1, respectively, and functionally validated their regulatory roles via site-directed mutagenesis. Functional assays revealed that phosphomimetic mutations, Cre1S204D and Ace1S145D, enhanced the repressive functions of Cre1 and Ace1, significantly reducing cellulase activity and down-regulating cellulase/hemicellulase gene transcription. In contrast, dephosphorylation-mimicking mutations, Cre1S204A and Ace1S145A, relieved carbon catabolite repression and promoted cellulase biosynthesis. Notably, the Cre1S204A/Ace1S145A double mutant exhibited synergistically enhanced cellulase activity (3.7-fold higher pNPCase activity than the parental strain), with significantly upregulated cellulase gene expression and increased extracellular cellobiohydrolase 1 accumulation. Mechanistically, Cre1 and Ace1 may regulate the cellulase gene expression in a phosphorylation-dependent manner. Overall, our findings identify Cre1 S204 and Ace1 S145 as functional phosphorylation sites that negatively regulate cellulase transcription, expand the post-translational regulatory network of fungal cellulase expression, and provide precise genetic targets for the rational design of high-efficiency fungi for lignocellulose bioconversion.

Abstract Drug–target interaction prediction and binding affinity prediction are two key tasks in drug discovery and drug repurposing. Although deep learning methods have made significant progress, existing models typically rely on global representations of drugs and proteins, making it difficult to adequately model fine-grained interactions between their local units. Fixed multimodal fusion strategies also struggle to dynamically adjust the contributions of different modalities for different drug–target combinations. To address these issues, we propose DQHTFI, a fine-grained interaction prediction framework for drug–target interaction classification and binding affinity regression. DQHTFI employs BRICS fragments and Pfam functional domains as the basic interaction units and jointly learns semantic and structural representations. We design a dynamic-query hypergraph Transformer framework in which hyperedges are constructed among the multimodal features of fragment–domain pairs. Dynamic queries are generated from the cross-conditioned features of fragment–domain pairs to adaptively adjust the contribution of each modality, thereby modeling higher-order interactions between local units. Our proposed model achieves competitive results on multiple benchmark datasets.

Peptides combine the favorable pharmacokinetics of small molecules with the high specificity of biologics, making them promising therapeutics. Incorporating non-canonical amino acids (ncAAs) further enhances drug-like properties, yet modeling remains challenging due to chemically modified residues and combinatorial sequence diversity. Here, we introduce SinCAA, a similarity-enhanced pretraining framework specifically designed to encode ncAAs. The framework is built on the principle that amino acids with similar 3D conformations induce minimal perturbations to peptide properties. It jointly optimizes two complementary self-supervised tasks: contrastive learning guided by a conformational similarity metric to capture functional relationships among ncAAs, and masked node reconstruction to encode the unique chemical identity of each ncAA. Built on a graph transformer backbone, this dual "relationship-identity" supervision enables SinCAA to learn robust atomic representations that generalize from individual ncAA building blocks to full-length peptides. SinCAA exhibits strong zero-shot performance in peptide property prediction and consistently outperforms state-of-the-art pretrained models across diverse benchmarks. This framework provides an efficient and interpretable approach for in silico prediction and ranking of ncAA-containing peptides, accelerating candidate screening in therapeutic peptide discovery.

Wheat, the most widely cultivated crop globally, includes multiple species with varying ploidy levels 1 . Globally cultivated species such as pasta and bread wheat belong to the Emmer lineage. In contrast, the Timopheevii lineage is of negligible agricultural significance 2 , yet it harbours genetic diversity of potential value for wheat improvement. Here, we present genomic resources for the Timopheevii wheat lineage. Our analyses uncovered extensive homoeologous exchange and genome instability in zhukovsky’s wheat, the hexaploid representative of the Timopheevii lineage. Population genomic analyses reveal a mosaic-like haplotype composition of domesticated tetraploid timopheev’s wheat, suggesting it was once cultivated across a broader geographic range. This is further supported by the discovery of ‘fossil’ haplotype segments in bread wheat landraces, likely derived from now-extinct domesticated timopheev’s wheat populations. By examining the Timopheevii lineage, we provide new insights into the evolutionary dynamics of wheat, discovering previously underexplored and potentially useful genetic diversity for wheat improvement.

Aortic stenosis (AS) is a disease with a high mortality rate. With advances in deep learning, some studies that use deep learning to aid in the diagnosis of AS severity have emerged in recent years. However, most of these studies have relied on complex deep learning models with a large number of parameters. In addition, many of these studies trained models using a fully supervised method, which required a significant amount of labeled data while neglecting the potential value of unlabeled data. In this paper, we propose a lightweight semi-supervised learning model to perform two tasks: view and aortic stenosis severity classification. Our method achieved excellent classification performance with a small number of parameters and effective utilization of large amounts of unlabeled data during training. We evaluated our method on the publicly available TMED-2 dataset. Our method achieved 97.8% balanced accuracy in view classification, 82.2% balanced accuracy in image-level AS severity classification, and 76.7% balanced accuracy in patient-level AS severity classification, outperforming existing methods. The implementation of our approach is publicly available at: https://github.com/849131302/Lightweight-Semi-supervised-Learning-Model-for-View-and-Aortic-Stenosis-Severity-Classification.⋆ This work is supported by National Key Research and Development Program of China (No.2024YFE0213800), Key Research and Development Plan Project of Heilongjiang Province (GA23C007).

Accurate prediction of protein-ligand binding affinity (PLA) is essential for efficient drug screening. However, existing methods often inadequately model multimodal molecular interactions and their combined effects on binding affinity. To address this challenge, this study introduces AtomBind, a multimodal information-driven heterogeneous graph neural network framework. First, an atomic-level heterogeneous graph is constructed, integrating sequence information, 3D structure, geometric constraints, and molecular representations generated by pretrained protein and chemical language models. Second, the Intra Encoder utilizes a variational graph autoencoder and equivariant graph neural network architecture to capture complex topological relationships at the molecular scale, addressing short-range atomic dependencies and dynamic structural variations. Finally, the Inter Encoder incorporates graph diffusion convolution and graph Transformer architectures, facilitating cross-molecular information transfer between protein pockets and ligands while simultaneously capturing both global and local intermolecular interaction features. Comparative experiments demonstrate that AtomBind exhibits superior predictive consistency and smaller errors compared to other models on two test sets. Ablation studies and pretrained language model analysis further validate the efficiency and robustness of AtomBind in multimodal information integration and affinity prediction. Additionally, the model demonstrates strong generalization capabilities and broad practical application potential in analyzing protein pockets, intermolecular interactions, and interpretability analysis.