Causal models and prediction in cell line perturbation experiments.

Area of Science:

• Computational biology
• Systems biology
• Pharmacology

Background:

• Cell line perturbation experiments are costly, limiting in vitro testing.
• Predicting cellular responses to novel perturbations requires robust computational models.
• Existing models face challenges in extrapolating to untested perturbations.

Purpose of the Study:

• To propose causal structural equations for modeling cellular responses to perturbations.
• To develop and compare Linear Regression (LR) and Causal Structure Regression (CSR) estimators.
• To provide a causal interpretation for the Cellbox model and compare its performance against LR and CSR.

Main Methods:

• Utilized causal structural equations to model perturbation effects.
• Derived LR and CSR estimators for response prediction.
• Analyzed the connection between CSR and the Cellbox ordinary differential equations (ODEs) model.
• Compared LR, CSR, and Cellbox performance through simulations and on a melanoma cell line dataset.

Main Results:

• CSR can predict effects of previously untested perturbations, unlike standard LR.
• An analytic connection was established between CSR and the Cellbox model, offering a causal perspective.
• In simulations, LR and CSR/Cellbox demonstrated distinct strengths and weaknesses.
• On the melanoma dataset, LR achieved performance comparable to or slightly exceeding Cellbox.

Conclusions:

• Causal modeling approaches like CSR offer advantages for predicting responses to novel perturbations.
• The Cellbox model can be causally interpreted through its connection to CSR.
• Linear Regression remains a strong baseline for perturbation response prediction, performing competitively with more complex models like Cellbox.