Cell： Using AI Agents to Drive Biomedical Discovery
Compilation Shu Wang Cong
Editor: Wang Duoyu
Typesetting Shu Shui Chengwen
The long-term vision for artificial intelligence (AI) is to develop AI systems that can make major scientific discoveries, learn autonomously, and acquire knowledge autonomously.
虽然“AI 科学家”（AI scientist）这一概念还只是一种理想化的愿景，但基于智能体（agent）的AI技术的发展，为开发能够进行对话、具备反思学习和推理能力的AI智能体（AI agent）铺平了道路，这些AI智能体能够协调大型语言模型（LLM）、机器学习（ML）工具、实验平台，甚至是它们的组合。
近日，哈佛大学医学院 Marinka Zitnik实验室（高尚华博士为第一作者）在 Cell 期刊发表了题为：Empowering biomedical discovery with AI agents 的文章，系统介绍了利用 AI 智能体（AI agent）推动生物医学发现。
该文章提出了这样一个观点——“AI 科学家”可以被由人类、大语言模型、机器学习模型及其他工具（例如实验平台）组成的复合AI系统所支持的 AI 智能体（AI agent）所实现。
我们将“AI 科学家”设想为具备批判性学习和推理能力的系统，它们能够通过协作智能体整合 AI 模型和生物医学工具，并与实验平台相结合，从而为生物医学研究提供助力。
生物医学 AI 智能体不是将人类排除在发现过程之外，而是将人类的创造力和专业知识与 AI 分析大量数据、探索假设空间和执行重复性任务的能力相结合。AI 智能体将能够熟练完成各种任务，规划发现工作流程并进行自我评估，以识别并解决其知识中的差距。这些智能体使用大语言模型（LLM）和生成式模型来实现结构化记忆以进行持续学习，并使用机器学习工具将科学知识、生物学原理和理论融入其中。AI 智能体可以影响从虚拟细胞模拟、可编程的表型控制到细胞回路设计以及开发新疗法等各个领域。
The first author, Dr. Gao Shanghua, received his doctorate from Nankai University and is now a postdoctoral researcher at Harvard Medical School
The complexity of biology requires the flexibility to break down complex problems into executable tasks. AI agents can break down problems into manageable sub-tasks, which are then solved by AI agents with specific functions and integrate scientific knowledge. In the near future, AI agents can accelerate the discovery process by making workflows faster and more resource-efficient.
Leveraging AI agents to advance biomedical research
The continuous development and application of data-driven models
AI agents in the biomedical field – from LLM-based AI agents to multi-agent systems that integrate AI models, tools, and physical devices
Level of autonomy of AI agents
Cases of autonomy levels for AI agents in genetics, cell biology, and chemical biology
Core terms of machine learning
Key modules in AI agents: perception, interaction, reasoning, and memory modules
Components of a biomedical AI agent
The challenges of AI agents in biomedical discovery
AI agents can improve the efficiency of routine tasks, automate repetitive processes, and analyze large datasets to outperform current human-driven efforts in scale and accuracy. This automation allows for continuous, high-throughput research that human researchers cannot conduct individually at the same scale or speed.
Looking to the future, AI agents can provide insights beyond what traditional machine learning can achieve by making predictions across temporal and spatial scales before obtaining experimental measurements at both temporal and spatial scales. Ultimately, AI agents may help uncover new patterns of behavior in biological systems.
Original link:
https://www.cell.com/cell/fulltext/S0092-8674 (24) 01070-5
Read the original text