This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
Xiangru Tang, Zhuoyun Yu, Jiapeng Chen, Yan Cui, Daniel Shao, Weixu Wang, Fang Wu, Yuchen Zhuang, Wenqi Shi, Zhi Huang, Arman Cohan, Xihong Lin, Fabian Theis, Smita Krishnaswamy, Mark Gerstein
Outline
CellForge is a virtual cell modeling system that aims to predict quantitative responses to various stimuli through the convergence of artificial intelligence and biology. To overcome challenges such as the complexity of biological systems, heterogeneity of data types, and the need for expertise across various fields, CellForge leverages a multi-agent framework to directly transform provided biological datasets and research objectives into optimized virtual cell models. Using only single-cell multi-omics data and task descriptions, CellForge generates optimized model architectures and executable code for virtual cell model training and inference. It consists of three core modules: Task Analysis for dataset feature analysis and literature search; Method Design for developing optimal modeling strategies through collaboration among expert agents; and Experiment Execution for automated code generation. Agents in the Method Design module, comprised of experts and mediators with different perspectives, collaborate until consensus is reached. Experiments on single-cell perturbation prediction using six diverse datasets (including gene knockout, drug treatment, and cytokine stimulation) demonstrate CellForge's superior performance compared to existing state-of-the-art methods. This demonstrates that iterative interactions among LLM agents with different perspectives yield superior solutions than directly addressing modeling tasks.
Takeaways, Limitations
•
Takeaways:
◦
We demonstrate that CellForge, a multi-agent-based virtual cell modeling system, can automatically generate optimized virtual cell models from single-cell multi-omics data.
◦
Achieves superior performance over existing state-of-the-art methods on a variety of datasets and tasks.
◦
Demonstrating the utility of a modeling approach through collaboration between LLM agents from different perspectives.
