Cell systems最新文献

Deriving principles governing cell biology from single-cell measurements across modalities, called multimodal modeling, can advance our understanding of cellular states in health and disease. Realizing the full potential of multimodal models requires learning generalizable representations across data types, diseases, and biological contexts. This perspective examines the potential of compositional AI as a modular design approach for constructing multimodal foundation models that unify biological modalities-such as chromatin accessibility, protein abundance, spatial transcriptomics, microscopy imaging, and textual annotations-into cohesive representations of cellular behavior. We present key deep learning modeling approaches, along with transformer-based attention strategies to implement them, while addressing challenges posed by limited data availability and structural differences between modality representations. We also discuss how to connect and align partially overlapping multimodal measurements to build a comprehensive representation space. By synthesizing these technical advancements, we chart a path toward agentic virtual cell models, offering insights into opportunities, limitations, and future directions for leveraging multimodal AI to decode the complexity of cellular systems.

Most diseases are not caused by large-effect single factors but by the cumulative impact of small, context-dependent perturbations arising from genetic variants, personal behavior, or environmental exposures, a phenomenon we term the "long tail" of biology. Early disease signals often differ from late-stage biomarkers and evolve across demographic, lifestyle, and environmental contexts. Shifting medicine from reactive treatment to proactive health requires detecting and interpreting these signals. This requires longitudinal, multimodal data collection; non-invasive, scalable biosensing platforms; new technologies for interrogating biological complexity; and AI models capable of contextual, mechanistic reasoning. We propose an "N-of-1 analyzer" framework to track divergence from personal baselines across analytes, relationships, networks, and trajectories, interpreted through digital-twin simulations and knowledge-grounded foundational models. This framework enables early, individualized insights into disease risk and system decline, offering a path toward scalable precision prevention. Regulatory innovations will have to evolve, embracing complexity instead of reducing it to the mean.

Synthetic biology aims to achieve predictable, programmable control over living systems by designing and engineering biological components and functions. Over the past 25 years, the field has advanced from foundational molecular tools to increasingly complex systems-level architectures. A new inflection point has emerged with the integration of generative artificial intelligence (AI), catalyzing a fundamental shift in how biological design is conceived and executed. Generative AI now enables the data-driven creation of novel designs with predictable functionality and context-aware precision. Here, we examine the convergence of synthetic biology and generative AI, highlighting key innovations at this emerging frontier of deep generative design across biological parts and systems. We discuss how design frameworks have evolved and outline the opportunities and challenges that lie ahead, spanning biomolecular elements, genetic circuits, and genomes. Finally, we propose a roadmap for how generative AI can unlock a new era of predictable, programmable synthetic biological systems.

Advancements in single-cell technologies and deep sequencing have expanded the B cell repertoire available for antibody discovery. However, selecting the highest-affinity antibodies from many sequences remains challenging, reflecting our incomplete understanding of the mechanisms sustaining affinity maturation and associated molecular markers. Here, we generated datasets of antigen-specific B cells after mouse immunization and reanalyzed public data to identify "High Signature" (HS), a transcriptomic signature predictive of high-affinity antibodies. HS was derived through differential expression analyses and machine learning by integrating antibody sequences, gene expression, and affinity measurements of expressed antibodies. HS enabled sub-nanomolar-affinity antibody selection without prior sequence analysis in de novo immunization campaigns. HS-expressing B cells were 3 times more likely to yield high-affinity antibodies than randomly picked cells. HS demonstrated translatability to two human PBMC datasets from COVID patients, resulting in enriched high-affinity antibody selection, highlighting its antibody discovery potential across species. A record of this paper's transparent peer review process is included in the supplemental information.

Biomedical research requires quantitative rigor, i.e., numeracy, a facility with numbers. The last decade has seen the broad adoption of statistical tools ("Numeracy 1.0"). To drive science forward, the expertise to quantitatively evaluate hypotheses and insights also needs to be broadly adopted ("Numeracy 2.0"). Systems biologists will be at the forefront of the transformation.

Are there general, systems-level principles guiding the evolution and design of natural or artificial sensory and signaling networks? Here, we argue that the signal transduction networks in living cells display important similarities in their organization and dynamical responses to both synaptic networks of brain cells and recent architectures of artificial neural networks. We propose that the key property of all of these networks-organization into multiple layers with hierarchically distributed timescales-is not accidental but rather reflects optimal processing of complex signaling and sensory inputs. We term this the hierarchical timescale hypothesis. We propose that the convergent evolution toward multi-step processing with "decreasing bandwidth" can also explain multiple properties of signaling networks, such as how a single input can control diverse outputs on different timescales and how noise and delay accumulation can be gracefully handled by the network.

Integrating single-cell or spatial transcriptomic and epigenomic data enables scrutinizing the transcriptional regulatory mechanisms controlling cell fate. Current integration methods usually align multi-omics data into a shared latent space but fail to reveal the underlying connections between genes and regulatory elements. The correlation- or regression-based regulatory inference methods cannot dissect different transcriptional regulation codes for cells under different spatial and temporal states. To address both problems, we develop a feature-guided optimal transport (FGOT) method, which simultaneously uncovers cellular heterogeneity and their associated transcriptional regulatory links. FGOT also provides post hoc interpretability for existing integration methods. FGOT is applicable for paired/unpaired single-cell multi-omics data and paired spatial multi-omics data. Benchmarking and validating via histone modification data or three-dimensional (3D) genomics data show good robustness and accuracy in integration and inference of regulatory links. The method allows systematic screening of cell-state and spatial-location-specific regulatory elements in diseases at the single-cell level. A record of this paper's transparent peer review process is included in the supplemental information.

Pathogens have evolved complex strategies that exploit the unique intracellular niches of organelles to establish a favorable replication environment that promotes infection and associated diseases. Defining how pathogens remodel organelle structures and compositions to redirect their functions is a major goal in cell biology. Recent technological advancements now enable structural characterizations of remodeled organelles in exquisite detail, as well as quantitative mapping of relocalized protein constituents and suborganellar interacting proteins. We describe emerging advances in complementary approaches for spatially and temporally profiling organelle rearrangements dictated by pathogen infection, with a focus on state-of-the-art microscopy, quantitative proteomics, and the integration of computational developments during virus infection. We examine the organellar resolutions and subcellular scales of these methodologies and recent applications during viral infections. We discuss how existing biochemical and bioinformatic tools can be integrated for systems-level mapping of organelle remodeling dynamics to dissect structure-function relationships of rewired organelles induced by microbes.