1763-Autonomous Agents for Virtual-Cell Modeling

Single-cell foundation models are transforming computational biology by leveraging massive multi-omic and spatial atlases to learn transferable representations of cellular states. These frameworks utilize diverse tokenization strategies and neural network backbones, such as transformers and large language models, to process sparse and high-dimensional biological data. Despite their growing scale, current architectures face significant open challenges regarding zero-shot generalization, perturbation modeling fidelity, and evaluation standards. Nevertheless, these models are increasingly shifting from descriptive feature extractors to experiment-guiding systems capable of generating testable biological hypotheses and simulating cellular dynamics.References:Jiang S, Fang Z, Zhang Y, et al. The Landscape of Single-Cell Foundation Models: Design Principles, Applications, and Open Challenges[J]. 2026.前往小宇宙评论区与主播互动

The paper introduces CellForge, an innovative framework that utilizes coordinated teams of autonomous AI agents to completely automate the design and implementation of virtual-cell models. By treating neural network architectures as scientific hypotheses, the system processes diverse single-cell omics data, analyzes biological literature through a critic-mediated discussion protocol, and generates competitive perturbation-prediction models without human design intervention. Through rigorous evaluations on genetic, chemical, and cytokine datasets, the multi-agent system consistently outperforms traditional baselines and produces research plans favored by blinded human experts.References:Gerstein M, Tang X, Chen J, et al. AI Scientists for Building Virtual-Cell Models[J]. 2026.前往小宇宙评论区与主播互动

The paper introduce trajectory replay, an innovative framework designed to update continuous-time generative models with sequentially arriving data without suffering from catastrophic forgetting. By replacing expensive sample-level generative replay with velocity-field regression along the historical model's own trajectories, the method halves the expected simulation cost and enhances robustness against discretization errors. Mathematically, the authors demonstrate that the optimum of this approach serves as a density-weighted mixture of historical and incoming probability paths, which effectively recovers a weighted forward KL divergence. Furthermore, the integration of a matching rule for the mean velocity enables efficient few-step sampling across the flow map. Extensive evaluations validate that trajectory replay outperforms traditional replay and distillation baselines on class-incremental image generation tasks. Finally, empirical assessments on biological case studies confirm its efficacy in extending generative perturbation models to new experimental contexts and modalities while successfully preserving treatment-effect predictions.References:Palma A, Jones C, Cinquin T, et al. Trajectory Replay: Continuous-Time Mixture Transport for Lifelong Generative Models[J]. arXiv preprint arXiv:2610.26749, 2026.前往小宇宙评论区与主播互动

The Chem-PerturBridge compendium integrates over 1.25 million transcriptomic samples across eight assay types and 37k compounds to address the fragmentation of public perturbation resources. This harmonized multi-dataset resource standardizes identifiers, metadata, and replicates to evaluate signature agreement and enable model pretraining. The authors find that cross-dataset logFC correlation is generally weak and frequently falls below same-context different-compound baselines. In contrast, logFC direction agreement is considerably more stable and usually exceeds these baseline values. Furthermore, pooled representation learning on Chem-PerturBridge improves prediction performance on held-out compounds compared to L1000-only embeddings and molecular fingerprints. Ultimately, the resource provides a unified diagnostic and modeling framework despite assay and batch discrepancies.References:Szałata A, Novitskaia O, Shulman M, et al. Chem-PerturBridge: a harmonized compendium of small molecule perturbation transcriptomic effects[J]. arXiv preprint arXiv:2605.31522, 2026.前往小宇宙评论区与主播互动

Recent scientific investigations utilize spatial transcriptomics to examine how chemokine networks establish distinct structural landscapes within the tumor microenvironment across multiple solid tumor models. The findings demonstrate that specific anatomical compartments maintain conserved and unique chemokine expression patterns, which coordinate the local recruitment, positioning, and retention of immune cells. In particular, perivascular dendritic cells produce specialized signaling molecules that trap lymphocytes within the stroma, thereby restricting their infiltration into the tumor core. Ablating these key populations forces immune cells to migrate inward, although this disruption ultimately diminishes overall anti-tumor immunity and accelerates tumor growth.References:Altenburger L M, Patil A, Jobst J, et al. Chemokine Landscapes of the Tumor Microenvironment[J]. bioRxiv, 2026: 2026.08. 08.743563.前往小宇宙评论区与主播互动

The paper introduce SQUINT, a graph vector-quantized variational autoencoder designed to process spatially-resolved transcriptomics data by simultaneously learning distinct cell and niche codebooks. This framework utilizes residual vector quantization to capture both cell-intrinsic identities and broader microenvironmental contexts in a hierarchical discrete token format. By incorporating domain-specific training objectives such as adjacency, contrastive, and batch-correction losses, the model successfully merges spatial structural information across multiple tissue sections and assays. Extensive benchmarking across diverse datasets demonstrates that SQUINT either outperforms or remains competitive with leading continuous representation and single-cell foundation models on tasks ranging from cellular identification to cross-section integration. Furthermore, the resulting discrete vocabulary facilitates advanced biological applications, including gene-expression imputation in hidden regions and streamlined query-to-reference mapping for complex tissue atlases.References:Birk S, Merchant A, Vahidi A, et al. Learning Discrete Cell and Niche Codes from Spatial Transcriptomics Using Dual Residual Vector Quantization[J]. bioRxiv, 2026: 2026.08. 07.743490.前往小宇宙评论区与主播互动

The single-cell Atlas Toolbox (scAtlasTb) is a modular, Python-based framework designed to streamline and standardize the iterative, complex process of constructing large-scale single-cell reference atlases. Built using Snakemake and integrated with the scverse ecosystem, the toolbox provides a robust infrastructure for data preparation, quality control, workflow management, and downstream analysis. To address the historical lack of uniform standards in atlas building, scAtlasTb incorporates quantitative benchmarking metrics—such as Moran's I for label-free biological conservation and graph dissimilarity scores—to systematically evaluate methodological choices. Through case studies on the Human Lung Cell Atlas, Human Retina Cell Atlas, and Human Cell Atlas Reference for PBMCs, the authors demonstrate how this framework optimizes lineage-specific quality control, batch variable definition, and feature selection strategies. Ultimately, scAtlasTb transforms reference atlas creation into a transparent, reproducible, and evidence-driven procedure capable of scaling to millions of cells while preserving subtle biological signals.References:Mueller M F, Cujba A M, Romanovskaia D, et al. Building optimized single-cell reference atlases with scAtlasTb[J]. bioRxiv, 2026: 2026.07. 30.741695.前往小宇宙评论区与主播互动

The paper introduces TERRA, a spatial foundation model pretrained on 112 million human cells using a joint-embedding predictive architecture to analyze imaging-based spatial transcriptomics. By tokenizing each cell alongside its local environment, the model simultaneously generates gene-, cell-, and neighborhood-level embeddings from a single pretrained backbone without requiring task-specific fine-tuning. This multi-scale framework supports zero-shot tissue analysis, cross-platform transfer, and in silico gene perturbations that predict cellular and microenvironmental responses to injury or therapeutic intervention. Furthermore, the sources demonstrate TERRA's practical utility by uncovering a developmental precursor of islet microvasculature during pancreas organogenesis, predicting a checkpoint-blockade injury program in the human kidney, and cataloging conserved macrophage niche archetypes across multiple organs and disease states.References:Birk S, Sanian M V, Vahidi A, et al. Multi-scale modeling of human tissues from spatial transcriptomics with TERRA[J]. BioRxiv, 2026: 2026.07. 29.741565.前往小宇宙评论区与主播互动

The paper introduces the omnigenic neural network (OGM), a biologically structured architecture designed for human genome-scale phenotype prediction using the Gene Ontology (GO) as an inductive bias. By integrating individual-level genotypes, sex, and genetic principal components into a hierarchical gene and ontology layer, the model effectively captures non-additive interactions, such as epistasis, and accommodates multimodal data like metabolomics. Evaluated across large cohorts including the UK Biobank and the All of Us cohort, the OGM matches or outperforms traditional polygenic scores (PGS) and baseline models, especially in multilabel predictions and cross-ancestry transferability. Furthermore, the architecture provides systems-level interpretability and captures eQTL structures through methods like Integrated Gradients and Integrated Hessians, establishing a flexible framework for advanced genomic research.References:Upmeier zu Belzen J, Arnoldt L, Hollmann N, et al. An interpretable omnigenic neural network architecture for the human genome[J]. medRxiv, 2026: 2026.07. 28.26359187.前往小宇宙评论区与主播互动

Researchers created a multimodal cellular and spatial atlas comparing Broca’s area and the motor cortex across human development from infancy to adulthood to understand the biological foundations of language. By combining single-nucleus RNA sequencing, chromatin accessibility profiling, and spatial transcriptomics, the investigation mapped distinct cell types, gene expression programs, and structural architectures within these critical brain regions. The findings revealed area-specific cellular signatures, including unique glial populations, interneuron subtypes, and an early molecular code driven by cadherin genes that facilitates precise neural communication. Additionally, the study linked specific cell types, such as somatostatin-positive interneurons, to increased vulnerability for neurodevelopmental disorders like autism spectrum disorder and attention-deficit/hyperactivity disorder. Finally, evolutionary analyses of differentially accessible chromatin regions suggested that genetic adaptations supporting human linguistic abilities accumulated gradually over millions of years following the divergence of human and chimpanzee lineages.References:Moriano J, Mukhtar T, Lee J T H, et al. A multimodal interrogation of Broca’s area in the pediatric and adult human brain[J]. bioRxiv, 2026: 2026.07. 22.739934.前往小宇宙评论区与主播互动