#35 · ML / AI methods

Models predict and audit cellular responses to perturbations across modalities

The dominant thread is computational modeling that links a perturbation or stress condition to a predicted cellular state and then interrogates that prediction. BaiZe is a multi-view conditional state-transition model that predicts post-perturbation transcriptomes from control states across genetic, chemical, and temporal contexts. The Stenotrophomonas digital twin pairs an LHS experimental design with FVA flux-width features, XGBoost with SHAP attribution, and causal analysis to locate bioprocess bottlenecks. The retrieval-augmented LLM framework is a distinct case in the same spirit: it generates and audits hypotheses for longitudinal Cell Painting morphology in RPE-1 cells under low-dose-rate radiation, with explicit citation-validity checks. What is new is the emphasis on verifying the model itself, whether through multi-view conditioning, causal diagnosis of metabolic flexibility and rigidification, or validity auditing of generated hypotheses.