I know what you think about papers in Cell. They are long. Mechanistic. “Complete.” They report deep biological insights. I know this because I hear it a lot from researchers. There is truth to it. Cell has published thousands of papers that fit these descriptors, and we will continue to. They have incredible value. But we don’t publish papers in Cell because they are long and mechanistic. We publish them because they are exciting. Cool. Awe-inspiring. We publish them because we think they provide something that our readers will want to learn about and will benefit from.

The virtual cell is envisioned as a multi-modal, multi-scale, and stateful world model that can simulate, perturb, and ultimately program cellular systems. This perspective outlines how this paradigm could transform cell biology from experimental trial-and-error toward simulation-driven discovery, counterfactual exploration, and rational intervention design.

Biological processes continuously adapt in response to intervention. Unlike static predictive models, biomedical world models support action-conditioned simulations for counterfactual reasoning, intervention design, and sequential planning. This perspective defines the key properties of biomedical world models and discusses the data, modeling, and evaluation challenges required to build them across biological and clinical scales.

Salaudeen et al. review common reliability failures in predictive and generative AI for healthcare, including erroneous model outputs, clinically unjustified performance differences, and deployment-time degradation. They examine why existing technical solutions fall short and argue for lifecycle-aware evaluation, continuous monitoring, and institutional governance.

Artificial intelligence (AI) is changing the way biomedical research is conducted by making it possible to integrate diverse types of biological data and tackle questions that were previously difficult to address. At the same time, its growing use also raises important questions about data quality, reliability, interpretability, and how computational predictions should be combined with biological knowledge and experimental validation. In this Voices piece, researchers from diverse disciplines share their perspectives on how AI is advancing their respective fields and highlight both the transformative opportunities and the key challenges that will shape the future of the life sciences.

scBaseCount is presently the largest public single-cell RNA-seq repository, containing over 502 million cells across 27 organisms and 75 tissues. An AI agent autonomously discovers, annotates, and uniformly reprocesses all 10× Genomics datasets in the SRA, creating a harmonized, continually updated resource for studying the diversity of cell biology and training AI models.

Tahoe-100M is an atlas of 100 million single-cell transcriptomes, capturing how 50 cancer cell lines respond to ∼1,100 drug-dose treatments. By pairing single-cell and molecular phenotypes at scale, the resource links drug mechanisms to cellular responses and provides an openly available substrate for training predictive models of cell behavior.

LongevityBench, Longevity-LLMs, and Longevity Claw evaluate the readiness of the state-of-the-art AI systems for spearheading aging research.

Agarwal et al. introduce MitoSpace, a self-supervised model trained on 4D lattice light-sheet microscopy data. The model resolves drug-induced mitochondrial phenotypes without labels, predicts membrane potential from morphology and dynamics, generalizes to unseen perturbations and lung organoids, and shows that representation quality improves progressively from 2D to 4D.

Modeling perturbation effects across large single-cell populations requires flexibility to capture heterogeneity. By training over sets of cells in a shared embedding space, State outperforms baselines at generalizing effects to new contexts. Cell-Eval, the framework used for this comparison, provides a comprehensive benchmark for future models.

The Virtual Cell Challenge returns in 2026 with a more demanding test of biological generalization: zero-shot prediction across multiple independent cellular contexts. Participants will build models to predict gene knockdown responses in a new Arc-generated dataset comprising unseen cell lines. The goal is to determine whether the best models can meaningfully close the gap between preclinical experimental predictions and human biology.

Recombining domains across an entire protein family, rather than relying on natural sequences shaped by evolution, generates synthetic “DESynR” transcription factors with enhanced function. DESynR AP-1 TFs reprogram CAR T cells into non-natural, therapeutically optimized states and outperform natural AP-1 factors in antitumor immunity.

(Cell 188, 6953–6970.e1–e8; November 26, 2025)

Drawing inspiration from Hilbert’s list of 23 mathematical problems that have focused the mathematical community’s attention for more than a century, we propose fifteen grand AI challenges to focus the biomedical community’s attention on critically relevant questions, most of which still lack effective predictive methodologies.

Adult intestinal size plasticity is driven not only by epithelial stem cells but also by remodeling of the surrounding visceral musculature. Sex- and reproduction-dependent muscle remodeling controls gut size and transit, revealing the intestinal muscle as an active regulator of adult organ adaptation.

Spatial multi-omics reveals macrophage states that constrain adaptive immunity in immunosuppressed non-melanoma skin cancer.

Beyond its motor dimension, ALS can also involve cognitive impairment, with affected domains varying across individuals. Rather than a single common disease process, this heterogeneity reflects regionally and cellularly distinct alterations in the prefrontal cortex that are only partially predicted by TDP-43 neuropathology.

An integrative cell and spatial atlas defines the organizing logic of the adult human cerebrovasculature, providing a framework for studying how vascular cell states support brain function and shape neurological disease vulnerability.

Arginine availability regulates arginyl tRNA levels and codon-dependent translation of MHC class I, tuning antigen presentation and shaping anti-viral and anti-tumor immunity.

Advances in AI have shown promise in protein structural information, but in this perspective, the authors argue that biology’s true complexity—emergent, evolved, and irreducibly messy—demands more than bigger models or more data and that unlocking a genuine “biological” AI requires anchoring machine learning in canonical, process-based “world models,” rather than chasing a universal theory built from parts alone.