Biology subjects

Ouyang, W.

Publications and source records attributed to Ouyang, W..

At least 19 recordsLinked to original sources

bioRxiv2026

NeuroPack: Advanced Learnable Packaging Toolkit Towards Ultra-High Compression Ratio for Volumetric Neuroimaging

High-throughput 3D microscopic imaging generates brain-volume datasets at terabyte to petabyte scale, posing significant storage challenges for large-scale neuroimaging data. In this study, we introduce NeuroPack, a hyperprior-driven 3D deep-learning-based packaging toolkit specifically designed for compressing 3D microscopic volumes at the mesoscale. NeuroPack encodes a compact hyperprior stream that conditions its entropy model on local voxel-wise variance, adaptively allocating bits to preserve mesoscale anatomical details for an optimal rate-distortion balance. To ensure reproducible deployment across heterogeneous microscopy datasets and computing environments, NeuroPack implements model integerization for the hyperprior side-information pathway, whose hyper-decoder predicts the Gaussian mean and scale parameters used by the arithmetic coder. By replacing platform-dependent floating-point hyper-decoding with deterministic fixed-point integer convolutions, NeuroPack produces bit-exact entropy models across CPUs, GPUs, local workstations, and cloud servers, thereby preventing decoding drift caused by floating-point rounding, fused-multiply-add ordering, or library-specific kernels. This integerized design preserves rate-distortion behavior while maintaining efficient encoding and decoding for large anisotropic 3D microscopy volumes. Our method enables high-fidelity microscopy reconstruction at 100- to 1000-fold compression across diverse modalities and species. Quantitatively, NeuroPack delivers the strongest overall rate-distortion performance in both BD-PSNR and BD-MSSSIM, with average relative gains of 73.4% and 50.8%, respectively, over the best competing codec. Importantly, NeuroPack preserves image fidelity required for downstream morphological reconstruction and analysis, as validated by simultaneous human annotation of original and decompressed confocal whole-brain zebrafish microscopy images. Together, NeuroPack provides a learning-based compression framework explicitly tailored for microscopy data, establishing a scalable storage solution for data-intensive neuroimaging and AI-empowered biological discovery.

bioengineering↗

bioRxiv2026

CENO: A Genome-Scale World Model for Evolutionary Sequence Interpretation and Programmable Regulatory Design

DNA encodes biological function across a continuum of sequence scales, from single-nucleotide and motif-level grammar to regulatory neighborhoods, chromatin-scale organization and evolutionary constraint. A useful model of genomes should therefore do more than classify short sequence windows: it should maintain nucleotide-resolution state over long contexts, score counterfactual mutations, condition on homologous sequence evidence and generate candidates that can be evaluated against structural or functional objectives. We define such a system operationally as a genomic world model: a general-purpose generative model of genome sequence space that unifies sequence understanding and sequence design through a shared state and likelihood interface. Here we introduce CENO, a family of long-context generative genomic world models designed to preserve local DNA grammar while extending usable context to regulatory and chromatin scales. CENO combines Mamba sequence-mixing layers, sparse attention layers and mixture-of-experts capacity in a single autoregressive backbone, and is trained at 300M, 600M and 1B parameter scales with a staged curriculum that progresses from 8k-token cross-domain genomic pretraining to 131k- and 1M-token whole-genome long-context continuation. We evaluate CENO under a unified world-model benchmark paradigm spanning retrieval, representation, counterfactual perturbation, reconstruction, evolutionary conditioning and design. CENO retains practical long-context inference and retrieves distal sequence in synthetic assays. In zero-shot long-context analyses, without task-specific fine-tuning, long-context continuation yields annotation- and chromatin-boundary-associated attention patterns and frozen-state representations that generalize across human cell types and mouse cell or tissue settings. To incorporate evolutionary information, we further post-train CENO on packed real multiple-sequence-alignment contexts and score variants by reference-mutant likelihood deltas, improving matched variant-effect prediction and producing evolutionary enrichment signals across species. Complementing these perturbation-based variant tests, we evaluate zero-shot long-sequence generation by partial-gene continuation, asking whether the model can recover withheld gene-scale sequence structure across eukaryotic, bacterial and archaeal species; recovery improves with model scale and later whole-genome long-context training. Finally, we use CENO as the backbone for a cell-type-specific enhancer design workflow in mouse cortex, coupling a CENO-based accessibility oracle with conditional supervised fine-tuning and oracle-guided reinforcement learning. Together, CENO provides a genome-scale sequence world-model framework for sequence interpretation, evolutionary reasoning, gene-scale reconstruction and programmable regulatory sequence generation.

genomics↗

bioRxiv2026

AI4Life Open Calls and Public Challenges: why, how, and what we have learned.

Within AI4Life, we ran three Open Calls and three Public Challenges (2023-2025), supporting 22 bioimage analysis projects from 151 applications and engaging 225 challenge participants, with the aim of applying FAIR deep learning in the life sciences. Our experience offers a view of the current state of bioimage analysis, the landscape of available tools, as well as the existing gaps between method developers, tool producers and potential users. It highlights that even after careful selection for AI-ready projects, most still require substantial effort to apply deep learning, and that the field still relies heavily on established, well-rounded methods to solve common problems. We come to the conclusion that for scientific AI in biology, the rate-limiting step is not methods and models but data, annotations, and shared infrastructure underneath them.

bioinformatics↗

bioRxiv2026

Structure and function of human NXPE1, a sialic acid O-acetyltransferase

Human genetic studies have identified defects in multiple mechanisms that predispose the risk of developing inflammatory bowel diseases (IBD), which include alterations in adaptive and innate immune responses, epithelial integrity and regulation of the intestinal mucus layer. Despite the importance of intestinal barrier integrity in the pathogenesis of IBD, essentially all current therapies modulate the immune responses. In this study, we determined the high resolution cryo-EM structure of human NXPE1, a IBD associated protein. Based on the structural homology, we identified NXPE1 as an O-acetyltransferase. Since NXPE1 is a pseudo gene in mouse, we generated knockout mouse model that lacked two of the mouse NXPE1 homologs, Nxpe2 and Nxpe4. The O-acetylation of sialic acid on red blood cells was abolished in the double knockout mice, confirming the sialic acid O-acetyltransferase function of NXPE1 family members. These findings underscore the potential of NXPE1 as a novel therapeutic target of the intestinal barrier functions for the treatment of IBD.

immunology↗

bioRxiv2026

BioEngine: scalable execution and adaptation of bioimage AI through agent-readable interfaces

Foundation models and curated repositories have transformed bioimage AI, yet most biologists cannot readily run, adapt, or extend them on available hardware. BioEngine lls this gap as the execution and adaptation layer between curated AI and scalable compute, deployable on a laptop, workstation, or cluster. Scientists then screen models, ne-tune from the browser, enable real-time smart microscopy, and deploy analysis applications, all by describing their goal to an AI agent.

bioinformatics↗

bioRxiv2026

Generative machine learning unlocks the first proteome-wide image of human cells

The spatial organization of proteins within cells governs virtually all cellular functions, yet current imaging can simultaneously visualize only tens of proteins, orders of magnitude below the thousands populating a single human cell. Here we present ProtiCelli, a deep generative model that simulates microscopy images for 12,800 human proteins from just three cellular landmark stains. Trained on 1.23 million Human Protein Atlas images, ProtiCelli outperforms existing methods in reconstruction accuracy and textural fidelity, and generalizes to unseen cell types and drug perturbations. Simulated images preserve hierarchical subcellular organization, recapitulate known protein protein interaction landscapes, and resolve compartment-specific functions of moonlighting proteins at single cell resolution. Remarkably, the model infers drug-induced changes in protein expression and localization from cell morphology alone, predicts cell cycle stage without dedicated markers, and enables unsupervised segmentation of subcellular compartments and spatial decomposition of gene sets into functional regions. We leverage ProtiCelli to generate Proteome2Cell, a dataset of 30.7 million simulated images spanning 2,400 virtual cells across 12 human cell lines, enabling hierarchical single-cell models that distinguish conserved from dynamic protein architectures. Integrated into the Human Protein Atlas, Proteome2Cell democratizes exploration of these virtual cells. By computationally bridging the experimental scalability gap, ProtiCelli establishes a foundation for spatial virtual cell modeling.

cell biology↗

bioRxiv2025

scMOBA: A conversational single-cell Multi-Omics Brain Agent across species

Single-cell and spatial multi-omics are revolutionizing our understanding of the complexity in the developmental, aging, and diseased brain, but integrating this knowledge across modalities and species remains challenging. To bridge this gap, we propose scMOBA, a conversational single-cell Multi-Omics Brain Agent established by a large language model, a gene encoder, and a cross-attention projector. The scMOBA was pre-trained on 130 million single-cell and spatial multi-omics data spanning the entire brain across diverse species, development, aging and diseases. The pre-training utilized a novel multi-omics Feature-Question-Answer (FQA) paradigm, enabling the model to generate biological answers from feature inputs and textual queries. This unique scheme facilitates superior zero-shot inference capabilities without requiring additional fine-tuning. We demonstrate that scMOBA achieves state-of-the-art performance in fine-grained cell type classification across different species and modalities, as well as in batch correction and multi-omics data integration. Furthermore, scMOBA significantly boosts the accuracy of kinds of critical downstream tasks, including cell-type specific aging clock construction and disease status prediction. Overall, scMOBA serves as a powerful scientific discovery engine for multi-omics brain research, advancing the precision prediction and early intervention of neurological aging and diseases.

bioinformatics↗

bioRxiv2025

Scalable integration and prediction of unpaired single-cell and spatial multi-omics via regularized disentanglement

Deciphering cellular states requires methods capable of integrating large-scale heterogeneous single-cell and spatial omics data. However, these data are typically unpaired due to destructive assays and further confounded by modality heterogeneity, technical noise, and immense scale. Here we present scMRDR, a scalable computational framework based on regularized disentangled representation learning for integrating fully unpaired single-cell and spatial multi-omics datasets. Built on a unified and structure-preserving architecture, scMRDR removes the need for pairing supervision while maintaining computational efficiency, enabling scaling to large datasets spanning multiple disparate omics modalities. Across diverse real-world benchmarks, scMRDR demonstrates strong performance in batch correction, modality alignment, and biological signal preservation. The framework further supports cross-modal translation across omics modalities and enables spatial coordinate imputation for non-spatial single-cell datasets using a reference atlas. The resulting spatial mapping allows spatially resolved analyses, including identification of spatially variable genes and characterization of epigenetic regulatory programs in their native tissue context. These capabilities position scMRDR as a scalable and versatile framework for large-scale multi-omics integration.

genomics↗

bioRxiv2025

Grassland Ecological Compensation Policy On Pastoral Production Efficiency--Evidence From Pastoral China

Grassland Ecological Compensation Policy(GECP) in the protection of grassland ecological environment at the same time also promote the transformation and upgrading of animal husbandry, pastoral production efficiency is one of the important indicators to evaluate the sustainable development of animal husbandry, but also has a direct impact on the economic benefits of the herders. Based on field questionnaire data from 468 herding households in the pastoral areas of Inner Mongolia Autonomous Region of China, this study used two-stage semi-parametric DEA model and mediated effect model to measure herding households pastoral production efficiency and analyze the effect and influence mechanism of GECP on herding households pastoral production efficiency, respectively.The results of the study show that: (1) the overall pastoral production efficiency of herding households in Inner Mongolia is low, with a mean value of 0.43, which is mainly caused by low pure technical efficiency. (2) GECP, the number of years of education of the household head, whether or not the head of the household receives technical training, the percentage of income from herding, the number of herding machines, the evaluation of the herding socialized services and whether or not to rent pasture have a significant positive effect on the herding efficiency of herding households. (3) the realization of appropriate scale management by herding households plays a part of the mediating effect in the process of affecting herding households pastoral production efficiency. The realization of moderate scale operation by herders plays a partial intermediary effect in the process of GECP affecting herders pastoral production efficiency, and the direct effect of this policy on the pastoral production efficiency of large-scale herders is significantly larger than that of small-scale herders. Based on the above conclusions, this paper puts forward the suggestions of improving the supporting policies of ecological bonus, demonstrating and popularizing the breeding technology, and guiding the moderate scale management.

ecology↗

bioRxiv2025

A PAM-CNN recognition approach to Muscle Fatigue of Special Operators by sEMG signals

Fatigue operation was the primary accident causes of large special equipment which could lead to serious economic loss and personnel injury or death, so the establishment of early fatigue warning methods was of great significance to reduce the accident rate. In view of the working environment and biomechanics properties of operators of special equipment, theoretical and experimental method was presented to analyze the fatigue mechanism and accumulate data samples, and then a new approach was proposed to recognize muscle fatigue by classifying these datasets. Firstly, in order to deduce fatigue mechanism, hidden Markov chain was adopted to build five-element theoretical model for studying interaction between machines and operators muscles during operation process, so that influence factors and tested points of muscles were approached; Surface electromyographic (sEMG) sensors were introduced to acquire muscle signals which were taken as sample data, and wavelet packet was adopted to extract features of these data. Subsequently, a convolutional neural network (CNN) classification model based on partial attention mechanism (PAM) was proposed to execute triple fatigue classification, moreover cross entropy-Adam algorithm was applied to train and optimize this model. Finally, large crane operators were taken as case studies to verify the proposed model. By comparison results indicated when the number of extracted features was 3 and the size of patch was [5,5], PAM-CNN model could approach the optimum solution of fatigue recognition, and the average accuracy could achieve 95.6% in test datasets, which showed a good potential in practical application.

bioengineering↗

bioRxiv2025

Venus-MAXWELL: Efficient Learning of Protein-Mutation Stability Landscapes using Protein Language Models

In-silico prediction of protein mutant stability, measured by the difference in Gibbs free energy change ({Delta}{Delta}G), is fundamental for protein engineering. Current sequence-to-label methods typically employ the two-stage pipeline: (i) encoding mutant sequences using neural networks (e.g., transformers), followed by (ii) the {Delta}{Delta}G regression from the latent representations. Although these methods have demonstrated promising performance, their dependence on specialized neural network encoders significantly increases the complexity. Additionally, the requirement to individually compute latent representations for each mutant site negatively impacts computational efficiency and poses the risk of overfitting. This work proposes the Venus-MO_SCPLOWAXWELLC_SCPLOW framework, which reformulates mutation {Delta}{Delta}G prediction as a sequence-to-landscape task. In Venus-MO_SCPLOWAXWELLC_SCPLOW, mutations of a protein and their corresponding {Delta}{Delta}G values are organized into a landscape matrix, allowing our framework to learn the {Delta}{Delta}G landscape of a protein with a single forward and backward pass during training. Besides, to facilitate future works, we also curated a large-scale {Delta}{Delta}G dataset with strict controls on data leakage and redundancy to ensure robust evaluation. Venus-MO_SCPLOWAXWELLC_SCPLOW is compatible with multiple protein language models and enables these models for accurate and efficient {Delta}{Delta}G prediction. For example, when integrated with the ESM-IF, Venus-MO_SCPLOWAXWELLC_SCPLOW achieves higher accuracy than ThermoMPNN with 10x faster in inference speed (despite having 50x more parameters than ThermoMPNN). The training codes, model weights, and datasets are publicly available at https://github.com/ai4protein/Venus-MAXWELL.

bioinformatics↗

bioRxiv2025

Cell shapes decode molecular phenotypes in image-basedspatial proteomics

The diversity of cellular and tissue structures can arise from a few basic cell shapes, which undergo various transformations based on biophysical constraints on cytoskeletal organization. While cellular geometry has been linked with selected biological processes such as polarity, signaling or morphogenesis, the orchestration of the whole proteome in association to cell shape is still poorly understood. In this study, using more than 1 million images of single cells stained for 11,998 proteins across 10 cell lines in the Human Protein Atlas database, we performed an integrated analysis of organelle, pathway and single protein levels in association to a 2D cellular shapespace. We found that cell and nuclear shapes across cell lines exist in a shared continuum. We also found that the subcellular organelle topology varies across cell lines, but remains robust within each cell lines shapespace. At the single protein level, we found that cells of different shapes in the same cell cycle phase might be preparing for different fates, and that many non-cell cycle proteins expressed shape-based abundance variation. Using the same coordinate framework defined by shape, we could analyze the distribution shift of protein spatial localization under drug perturbation.

bioengineering↗

bioRxiv2025

Unveiling Protein-DNA Interdependency: Harnessing Unified Multimodal Sequence Modeling, Understanding and Generation

The interdependent orchestration of proteins, coding DNA sequences (CDS), and adjacent regulatory DNA elements underpins biological functionality, where evolutionary constraints shape codon usage, transcriptional regulation, and structural stability. We present ProDMM, a multimodal framework that deciphers these systemic relationships through unified sequence modeling. Integrating a context-aware encoder with conditional generation, our architecture captures codependent patterns between protein motifs and cis-regulatory syntax. ProDMMs encoder demonstrates superior predictive capabilities in forecasting pathway-level phenotypes and protein expression profiles, outperforming conventional single-modality approaches. Its decoder generates synthetically optimized regulatory elements by resolving spatial dependencies between coding sequences and adjacent non-coding regions, enabling transcriptional fine-tuning unavailable to sequence-agnostic methods. By resolving how localized CDS-NCDS-protein interdependencies propagate to system-level phenotypes, ProDMM establishes a paradigm for holistic biological sequence engineering. This approach enables coordinated optimization of translation efficiency, enzymatic activity, and pathway flux, offering transformative potential for designing tailored biosynthesis systems and advancing sustainable bioproduction strategies.

bioinformatics↗

bioRxiv2025

scWGBS-GPT: A Foundation Model for Capturing Long-Range CpG Dependencies in Single-Cell Whole-Genome Bisulfite Sequencing to Enhance Epigenetic Analysis

Single-cell DNA methylomes are challenging to interpret because of sparse CpG coverage and the complexity of genome-wide sequences. We present scDNAm-GPT, a universal foundation model that uses context-aware CpG tokenization, a Mamba backbone, and cross-attention to capture both local and global DNA methylation patterns. Trained on over one million single cells from 35 human and mouse tissues, scDNAm-GPT enables accurate cell clustering, zero-shot prediction of CpG effects on gene expression, improved trajectory inference, and reference-free deconvolution of cell types from cell-free DNA. The model hierarchically learns regulatory features, and its attention maps highlight functionally relevant regions, demonstrating high biological interpretability. These results establish scDNAm-GPT as a scalable and generalizable framework for single-cell epigenomic analysis, offering new opportunities to dissect epigenetic regulation in health and disease. Code is available at GitHub (https://github.com/ChaoqiLiang/scDNAm-GPT).

bioinformatics↗

bioRxiv2024

UniSPAC: A Unified Segmentation Framework for Proofreading and Annotation in Connectomics

Reconstructing dense neuronal connections from volume electron microscopy (vEM) images is a critical challenge in neuroscience, driving the development of various automatic neuron segmentation methods1. Although current state-of-the-art automated segmentation methods can achieve high segmentation accuracy, they still require substantial manual proofreading and rely heavily on labeled datasets, which are often scarce, particularly for non-model organisms. Here, we introduce a Unified Segmentation framework for Proofreading and Annotation in Connectomics (UniSPAC) by providing the interactive segmentation model in 2D-level and the neuron tracing model in 3D-level. UniSPAC-2D allows users to correct its segmentation errors through point-based prompts, combining segmentation and proofreading in a single framework. UniSPAC-3D automatically traces neurons segmented by UniSPAC-2D across image slices, significantly reducing human involvement. Furthermore, UniSPAC-2D and UniSPAC-3D models can facilitate the semi-automatic generation of labeled data for new species, eliminating the need for external annotation tools. The fresh annotated data generated during proofreading in turn optimizes the interactive model through an online learning strategy, reducing the labeling effort for novel species over time. UniSPAC outperforms the start-of-the-art Segment Anything Model (SAM) in Drosophila segmentation, achieving 47x higher efficiency, and surpasses ACRLSD in cross-species segmentation on zebra finch data.

neuroscience↗

bioRxiv2024

Deep generative model for protein subcellular localization prediction

Protein sequence determines not only its structure but also its subcellular localization. Although a series of artificial intelligence models have been reported to predict protein subcellular localization, most of them provide only textual outputs. Here, we present deepGPS, a deep generative model for protein subcellular localization prediction. After trained with both protein primary sequences and protein subcellular localization fluorescence images, deepGPS shows the ability to predict cytoplasmic and nuclear localizations by reporting both textual labels and generative images as outputs. In addition, deepGPS shows potential to be further extended for other types of subcellular localization prediction, even with limited input data volumes for training. Finally, an openGPS website (https://bits.fudan.edu.cn/opengps) is constructed to provide a public and convenient platform for protein subcellular localization prediction with the scientific community.

bioinformatics↗

bioRxiv2024

Integration of the DD-genome reshapes gene transcription, chromatin architecture and metabolome of allohexaploid wheat leading to enhanced adaptability

The integration, through hybridization, of the DD genome into domesticated tetraploid wheat gave rise to allohexaploid wheat, the most cultivated wheat globally growing across diverse environmental conditions. However, the regulatory basis of this integration on increased environmental adaptability in allohexaploid remains largely unexplored. Here, we investigated the change of transcriptome, epigenome as well as the chromatin interactome, and metabolome in three independent polyploidization events representing DD genome integration. Our findings reveal that polyploidization events induce the activation of defense-related genes through comprehensive reorganization of epigenome and chromatin architecture. DD integration not only brings an additional gene copy but also activates the homoeologs existing in the A and B subgenomes through chromatin interactions. Furthermore, secondary metabolites represented by alkaloids and flavonoids that are crucial for environmental adaptation, are significantly enriched following polyploidization. Thus, hexaploid wheat exhibits enhanced tolerance to alkalinity, UV-B light stress and high salt conditions was observed. These results highlight the indispensable role of DD genome integration in the adaptability of allohexaploid wheat during its evolution.

plant biology↗

bioRxiv2024

Genomics-FM: Universal Foundation Model for Versatile andData-Efficient Functional Genomic Analysis

Artificial intelligence (AI) plays a crucial role in functional genomic analysis, offering great potential for comprehending biological phenomena such as heredity, development, diseases, and evolution. However, the development of AI models needs substantial labeled data, and these models are typically task-specific with limited generalizability to various applications. Here, we develop Genomics-FM, a genomic vocabulary driven foundation model that enables versatile and label-efficient functional genomic analysis. Specifically, Genomics-FM is first pretrained with ensemble genomic vocabulary on vast unlabelled data to learn comprehensive and generalizable representations and then finetuned with specific genomic vocabulary on limited labeled data to selectively activate and adapt the pretraining knowledge for specific tasks. We show that Genomics-FM significantly reduces the dependence on labeled data, and demonstrates the capability to outperform existing models across a comprehensive suite of tasks including genome annotation, epigenomic and expression profile prediction, and variant effect assessment. Remarkably, Genomics-FM even shows impressive zero-shot predictive capabilities across diverse species and tissues and exhibits noticeable adaptability to RNA-related tasks. With feasibility in data scarcity and even cross-domain biological scenarios, Genomics-FM will promote the broad application of AI and empower researchers to tackle previously insurmountable challenges, paving the way for groundbreaking research and discoveries.

bioinformatics↗