scAgents: A Multi-Agent Framework for Fully Autonomous End-to-End Single-Cell Perturbation Analysis

While AI agents for scientific discovery have proliferated, truly end-to-end autonomous solutions remain scarce, particularly in complex interdisciplinary fields like single-cell genomics. We introduce scAgents, a fully autonomous multi-agent framework that transforms raw single-cell data and task descriptions directly into optimized computational solutions. Given only a dataset and research objective as input, scAgents outputs both a novel model architecture and executable code for training and inference without human intervention. When evaluated on the scPerturb datasets and benchmarks, scAgents consistently outperforms task-specific state-of-the-art methods, achieving up to 49% reduction in prediction error compared to scGPT for gene knockouts and Pearson correlation increases of up to 20% in expression predictions versus ChemCPA for drug perturbations. scAgents' ability to succeed where existing foundation models struggle is particularly significant, adapting effectively to different data types (scRNA-seq, scATAC-seq, CITE-seq) and various perturbation categories with consistent performance across modalities. Our code and some scAgents-designed models are available at https://anonymous.4open.science/r/scAgents-2025-242E/.