PRAXIS-VirtualCell proposes a modular framework that organizes biological data, models, perturbations, and validation evidence, using biological contracts and evidence-aware execution to separate supported predictions from extrapolation and abstention across E. coli, S. cerevisiae, and human K562 cells.

It proposes a modular framework that organizes biological data, predictive models, perturbations, adapters, execution environments, and validation evidence so that virtual experiments are reproducible and auditable. Virtual cells have largely grown out of single-task predictive models, leaving heterogeneous data, models, and validation evidence without a unified organizational framework; this framework brings those elements into one organizing structure. At the abstract level, the text describes the framework's modular components and goals, without implementation details or quantitative evaluation.

The system supports cross-species tasks spanning Escherichia coli, Saccharomyces cerevisiae, and human K562 cells. It extends virtual-cell tasks from a single species or single task toward multiple species settings as the scope of a unified runtime. The abstract explicitly lists the three species as the supported scope, without reporting per-species performance numbers.

Through biological contracts and evidence-aware execution, it distinguishes supported predictions from extrapolation and abstention. Beyond producing predictions, it introduces a judgment about the degree of evidential support, allowing the system to flag extrapolation and choose abstention rather than only emitting predicted values. The abstract states the existence and role of this mechanism, without a formal definition of the contracts or an evaluation of abstention decisions.

With agentic orchestration, it automatically translates natural-language questions into traceable virtual experiments. An agent layer sits on top of the framework so that natural-language questions can directly drive virtual experiment workflows while retaining traceability. The abstract describes this capability, without providing accuracy figures or case counts for the natural-language-to-experiment translation.

The framework targets researchers and engineering teams who need reproducible, auditable virtual experiments, applies to modeling and perturbation tasks across Escherichia coli, Saccharomyces cerevisiae, and human K562 cells, and addresses settings where natural-language questions drive experiments. Its design goal is to bring data, models, perturbations, adapters, execution environments, and validation evidence into a unified runtime, and to use biological contracts and evidence-aware execution to distinguish supported predictions from extrapolation and abstention, thereby providing a runtime foundation for trustworthy and scalable virtual-cell systems.

The visible text is an abstract and does not include figures, experimental setup, or quantitative metrics, so the concrete form of the biological contracts, the actual behavior of abstention decisions, and prediction accuracy on cross-species tasks cannot be judged. Readers should still watch how evidence-aware execution triggers abstention, how independently reproducible the traceable experiments generated by agentic orchestration are, and how the framework connects with existing virtual-cell models.