Virtual Embryo Challenge

Embryogenesis is one of the most fundamental yet least understood processes in biology. A single fertilised cell gives rise to a complete organism through hierarchical spatiotemporal orchestration of gene regulation, cell fate transitions, tissue morphogenesis, and organ formation. When this process is disrupted, the consequences can be severe: congenital defects affect 1 in 33 newborns and remain a leading cause of infant mortality worldwide. Despite decades of progress, we still lack predictive and generative AI frameworks that can model when, where, and how embryonic development goes off course.

Recent advances in single-cell and spatial genomics have created unprecedented opportunities to model development at the whole-embryo scale, turning embryogenesis into a compelling frontier for the NeurIPS community. The problem presents a new frontier for machine learning: learning to predict, generate, and perturb complex biological systems across space, scale, and time. Large single-cell and spatial transcriptomics-based embryonic atlases across species provide rich molecular snapshots across developmental time, but they do not by themselves reveal how cell states transition, how local molecular changes propagate to tissue- and organ-level phenotypes, or how development responds to perturbations. This creates a challenging benchmark for models that must jointly learn spatial context, temporal dynamics, biological hierarchy, and perturbation effects.

The Virtual Embryo Challenge addresses this gap by establishing a standardised benchmark for predictive embryogenesis modelling. The competition will advance generative, causal, and agentic AI systems that model embryonic development across space, scale, and time under genetic perturbations. Grounded in multimodal whole-embryo and early cardiac development datasets spanning eleven developmental stages, the benchmark will evaluate models on spatial context, multiscale reasoning, and temporal dynamics. By releasing curated datasets, baselines, and evaluation tools, this competition will catalyse open innovation toward robust, interpretable, and generalisable virtual embryo models.

Ultimately, this challenge aims to catalyse the development of predictive digital twins of mammalian embryogenesis: models that can simulate how an embryo develops, predict when and where development goes awry, and eventually guide targeted interventions. If successful, this could help shift the study and treatment of congenital disease from reaction toward prediction — and ultimately, prevention.