Towards a Virtual Embryo: A new approach methodology for probabilistic assessment of neural tube closure defects

Abstract

Neural tube closure represents a critical morphogenetic event during early vertebrate embryogenesis. When genetic or chemical factors disrupt this process, embryos can develop severe birth defects such as spina bifida. To understand how these defects emerge, we built a computational agent-based model that functionalizes cell signals and biomechanical responses. This computational approach allows us to simulate how the neural tissue elevates, bends, and fuses. By altering gene expression within this system, the model can estimate the probability that specific defects will occur. These simulated defects stem from complex biological interactions. Translating these interactions into animal-free chemical safety assessments requires a robust foundation. To provide this biological foundation, we constructed an Adverse Outcome Pathway network. This network maps how early molecular events cascade into neural tube defects. Instead of relying on simple linear pathways, the network details how diverse mechanisms interact, including how morphogen gradients shift, how cytoskeletons adapt, and how tissues pattern themselves. Mapping these complex pathways highlights the importance of identifying the initial chemical interactions, known as Molecular Initiating Events. To identify these early triggers, we designed an automated artificial intelligence pipeline. The pipeline processes chemical structures as molecular graphs using a Knowledge-Guided Pre-training of Graph Transformer architecture. By evaluating these graphs, the resulting models predict potential chemicals that trigger molecular initiating events related to developmental toxicity more accurately than current QSAR-based methods. Automating this analytical step allows researchers to prioritize hazardous chemicals rapidly for further experimental testing. While early chemical interactions trigger toxicity, observing how these biological pathways shift over time requires dynamic experimental data. Because neurulation relies on strict timing, traditional single-endpoint assays often fail to capture these crucial temporal changes. We addressed this gap by defining a time-resolved transcriptomic reference for zebrafish neurulation. Exposing embryos to various chemicals enabled us to measure RNA sequences across five distinct morphological stages. The resulting temporal signatures reveal how specific toxicants perturb neural tube pathways and provide a baseline to evaluate dynamic biological responses. Evaluating these dynamic responses ultimately requires linking the molecular data back to physical outcomes. We achieved this integration by feeding the time-resolved zebrafish transcriptomic data into an updated version of our agent-based model. The updated computational model uses gene expression fold-changes to modulate pathway activity and simulate how closure dynamics change under chemical stress. This data-driven simulator translates complex molecular signals into probabilistic defect outputs. By bridging experimental transcriptomics and dynamic modeling, this workflow advances how we assess developmental toxicity and reduces our reliance on animal models to protect human health and the environment.