Auditing Retrieval-Augmented LLM Hypotheses for Longitudinal Cell Painting Morphology

This paper presents an evaluation-first, retrieval-augmented framework for interpreting longitudinal Cell Painting morphology. Weekly treated-control morphology deltas, perturbation neighbors, pathway context, and literature evidence are linked through stable evidence identifiers before an LLM generates hierarchical, provenance-preserving hypotheses. The study uses a 9-week RPE-1 cell time course across five ionizing-radiation dose rates, from 0.003 to 6.0 mGy/hr. Two audits test whether cited identifiers exist in the prompt and whether predicted biological processes are compatible with altered morphology features. The framework suggests metabolic reprogramming and proteostatic stress at lower dose rates, but lacks ground-truth mechanism labels.