Virtual cells: progress and challenges from mechanistic modeling to data-driven integration

Virtual cells aim to represent cellular states computationally and predict responses to genetic, chemical, environmental, temporal, and spatial conditions. For comparative purposes, this review uses state encoding (E), state transition (T), and observation decoding (D) as a conceptual abstraction spanning explicit mechanistic models, data-driven foundation and perturbation models, multimodal systems, and spatial and multiscale simulation. We examine model boundaries, data foundations, resource disclosure, out-of-distribution evaluation, causal evidence, and experimental validation. Independent benchmarks show that parameter count and training-set size do not automatically yield robust out-of-distribution generalization, while transcriptomic representations alone do not capture the full functional state of a cell. Trustworthy virtual cells will require perturbation-responsive transitions, cross-modal alignment, continuous-time dynamics, mechanistic constraints, calibrated uncertainty, independent experimental validation, and, where appropriate, closed-loop experimental feedback.