The paper introduces CellForge, an innovative agentic system designed for autonomously creating virtual cell models to predict cellular responses to various biological perturbations. This system employs a multi-agent framework, where specialized AI agents collaborate across three core modules: Task Analysis (characterizing datasets and retrieving literature), Method Design (developing optimized modeling strategies through expert discussion), and Experiment Execution (generating and validating executable code). CELLFORGE consistently outperforms existing methods in diverse perturbation prediction tasks, demonstrating significant improvements in accuracy and correlation across different data modalities. Its strength lies in its ability to adaptively design model architectures by integrating biological knowledge and iteratively refining solutions through graph-based expert discussions among agents, rather than relying on predefined heuristics.
References:
