Beyond the drug-centric view: Advancing AI virtual cell platforms for environmental perturbation modelling

Environmental exposures contribute substantially to global morbidity and mortality, with their bio¬logical effects shaped by factors such as dose, exposure duration, life-stage timing, and cumulative or sequential exposure patterns. With rapid advances in AI-based virtual cell technologies, current frameworks emphasize genetic and pharmacological perturbations, leaving environmental pertur¬bations insufficiently modelled. This imbalance is due to fundamental structural limitations in data availability, chemical space coverage, and representational design, as well as lacking stand¬ardized annotation of dose, timing, and mixtures, which are core determinants of environmental perturbations. This review argues that incorporating environmental perturbations as a foundational component of virtual cell development is necessary to extend these models toward environmental and public health applications. We review the evolution of these perturbation predictive models from mathematical models to deep learning and foundation model architectures, evaluate the status of virtual cell efforts across genomic, transcriptomic, proteomic, metabolomic and phenomic modal-ities, and show the systematic gaps that currently limit their applicability to environmental health. We outline key challenges, including data scarcity and bias, inadequate representation of environ¬mental perturbations, limited multimodal integration, and weak benchmarking practices. To address these issues, we propose the development of standardized environmental perturbation datasets, integrated and standalone multimodal architectures, uncertainty-aware evaluation metrics, and regulatory-aligned benchmarking frameworks, which are all geared toward advancing in silico perturbation and non-animal testing, promoting the 3Rs paradigm.