VIRTUAL CELLS AS CAUSAL WORLD MODELS: A PERSPECTIVE ON EVALUATION

ICLR 2026 Conference Submission14498 Authors

Keywords: AI virtual cells, causal world models, causal evaluation, intervention validity, counterfactual consistency, trajectory faithfulness, mechanistic alignment, causal benchmarks, perturbation datasets, evaluation metrics, predictive accuracy vs causality, model reliability, biological simulation, taxonomy of causal metrics, trustworthy AI in biology

TL;DR: AI virtual cells must move beyond prediction to causal evaluation, testing interventions and mechanisms to ensure they function as reliable models of biology.

Abstract: This perspective argues that evaluating AI virtual cells requires moving beyond predictive accuracy toward assessing their ability to function as causal world models of biology. Existing benchmarks emphasize fit to observed data, rewarding pattern matching but failing to test responses to interventions. We propose that trustworthy virtual cells require causal evaluation: metrics and benchmarks that assess intervention validity, counterfactual consistency, trajectory faithfulness, and mechanistic alignment. Our contribution is two-fold: (1) a survey of recent approaches to virtual cell modeling, and (2) a taxonomy of causal evaluation metrics mapped to available perturbation datasets and benchmarks. By outlining gaps and proposing unified causal benchmarks, we position causal evaluation as the key step toward making virtual cells reliable world models of biology.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 14498