Sam Cox, Automating Chemistry With LLM-Based Agents; Primer: Andrew White

Andrew White
Head of Science & Co-Founder at FutureHouse,
Associate Professor of Chemical Engineering at the University of Rochester

Primer: A progress report on the automation of science

The intellectual bottlenecks of science are growing with exponential growth in research paper counts, complexity of papers, and a concurrent decline in scientific productivity. The next major breakthroughs will increasingly rely on the automation of the stages of scientific discovery. We have made progress on this grand challenge using large language models augmented with access to tools, called scientific agents. We have used these agents for automating literature research, designing molecules, engineering enzymes, and bioinformatics analysis. A major component of this is evaluation against expert researchers in fair comparisons. We have succeeded in training these agents and exceeding human performance across multiple tasks. Finally, I will discuss preliminary work on scientific reasoning models - a new direction of artificial intelligence in the scientific domains.

Sam Cox
Member of Technical Staff at FutureHouse,
Ph.D. Candidate at the University of Rochester

Meeting: Automating Chemistry With LLM-Based Agents

Large Language Models (LLMs) have demonstrated strong performance in tasks across domains, but struggle with complex chemistry-related problems. Moreover, these models lack access to updated knowledge sources and software, limiting their usefulness in scientific applications. We use LLM-based scientific agents to address this gap by integrating expert-designed tools. ChemCrow is an LLM chemistry agent designed to accomplish tasks across organic synthesis, drug discovery, and materials design, including novel design. Similarly, MDCrow leverages molecular dynamics tools to automate protein simulation setup, execution, and analysis. While frontier LLMs generally perform better on chemistry tasks, we find that open-source models, when augmented with appropriate tools, can achieve comparable results.

For more information visit: .