Clarify the essentials before you begin.

An AI virtual cell uses data and computational models to represent cellular states and study responses to drugs, genetic changes or environmental conditions. Individual models cover selected cell types, modalities and interventions; they are not complete replicas of living cells.

Cell state prediction concerns current or future molecular, morphological and functional features. Drug response simulation additionally considers compounds, doses and exposure time to study changes following treatment. They may share representations but require different tasks and evaluation criteria.

Define the task first. Single-cell transcriptomic tasks typically require expression matrices, cell and sample metadata, controls, perturbation conditions, doses and time points. Imaging tasks require source images, acquisition settings and relevant labels. Suitability depends on data quality, coverage and usage rights.

Cellular predictions do not directly establish efficacy in humans. Pharmacokinetics, tissue context, immune responses and safety require additional evidence. Models help develop and prioritize hypotheses; key conclusions still require experimental and subsequent research validation.

Consider independent test performance, simple baselines, generalization across cellular backgrounds and biologically meaningful readouts. Separate training and test data according to the task to avoid leakage through shared donors, batches or similar conditions.

They depend on data quality, the task, compute needs and validation resources. Scope, deliverables, milestones, fees and acceptance criteria are confirmed in the project proposal; no fixed performance or timeline is assumed.