Explicit Mechanistic Operators in Predictive Virtual Cells: Evidence, Failure Modes, and a Minimum Falsification...

The term virtual cell now covers objects ranging from curated biochemical simulators to atlas-trained neural representations, yet the decisive empirical question is narrower: whether a model built to obey known biochemical rules predicts the effects of new drugs or genetic changes better than an otherwise identical model that lacks them. Formally, this Perspective asks whether a biologically specified mechanistic operator improves out-of-distribution prediction of single-cell perturbation responses beyond strong simple, relational, transport-based, and learned-dynamical comparators. Recent benchmarks show that rankings depend on partition, statistical unit, effect-size distribution, gene set, and metric, and that several deep and foundation-model approaches fail to exceed no-change, mean-effect, additive, or linear controls. Set-based, knowledge-graph, transport, and learned state-equation models can improve defined tasks. No identified study isolates the incremental value of a biologically specified operator against late fusion, a capacity-matched static model, a capacity-matched learned-dynamical operator, and topology-rewired controls under one locked evaluation. Six source categories are distinguished, with the operator category divided into biologically specified and learned-dynamical subclasses. A minimum falsification framework is specified for the unfolded protein response with a declared capacity ledger, endpoint times chosen by a prespecified Stage 1 rule, and a locked comparator set. The framework is prospective: no new data, model implementation, or performance result is reported, and every threshold is a design choice awaiting empirical test.