Meet DrugAgent: A Multi-Agent Framework for Automating Machine Learning in Drug Discovery

Abstraction: LLM multi-agent system automating end-to-end drug discovery ML pipelines

Key points:

• Developed by researchers from USC, Carnegie Mellon, and Rensselaer Polytechnic Institute; automates machine learning programming for drug discovery without requiring coding expertise from pharmaceutical scientists
• Two core components: LLM Instructor (identifies domain-specific requirements, builds tools) and LLM Planner (manages exploration and refinement of multiple ML approaches)
• In a PAMPA dataset ADMET (absorption, distribution, metabolism, excretion, toxicity) prediction case study, DrugAgent achieved F1 score of 0.92 using a random forest model selected over graph neural networks and ChemBERTa
• Outperforms ReAct (general-purpose LLM reasoning framework) which failed on domain-specific API calls and lacked self-debugging; DrugAgent completed the full pipeline without human intervention
• Dynamic idea space management generates multiple approaches at the start and iteratively prunes based on experimental outcomes

Connections: Drugagent · University Of Southern California · Carnegie Mellon University · AI Agents · Drug Discovery · Large Language Models · Multi Agent Systems

Source: https://www.marktechpost.com/2024/12/01/meet-drugagent-a-multi-agent-framework-for-automating-machine-learning-in-drug-discovery/