Carnegie Mellon University researchers have pioneered an artificially intelligent system, Coscientist, that can autonomously develop scientific research and experimentation. Published in the journal Nature, this non-organic intelligent system, developed by Assistant Professor Gabe Gomes and doctoral students Daniil Boiko and Robert MacKnight, is the first to design, plan, and execute a chemistry experiment autonomously.
Utilizing large language models (LLMs) like OpenAI’s GPT-4 and Anthropic’s Claude, Coscientist demonstrates an innovative approach to conducting research through a human-machine partnership.
Coscientist’s design enables it to perform various tasks, from planning chemical syntheses using public data to controlling liquid handling instruments and solving optimization problems by analyzing previously collected data. Its architecture consists of multiple modules, including web and documentation search, code execution, and experiment automation, coordinated by a central module called ‘Planner,’ a GPT-4 chat completion instance. This structure allows Coscientist to operate semi-autonomously, integrating multiple data sources and hardware modules for complex scientific tasks.
“We anticipate that intelligent agent systems for autonomous scientific experimentation will bring tremendous discoveries, unforeseen therapies, and new materials,” the research team wrote in the paper. “While we cannot predict what those discoveries will be, we hope to see a new way of conducting research given by the synergetic partnership between humans and machines.”
The system’s capabilities were tested across different tasks, demonstrating its ability to precisely plan and execute experiments. For instance, Coscientist outperformed other models like GPT-3.5 and Falcon 40B in synthesizing compounds, particularly complex ones like ibuprofen and nitroaniline. This highlighted the importance of using advanced LLMs for accurate and efficient experiment planning.
A key aspect of Coscientist is its ability to understand and utilize technical documentation, which has always been a challenge in integrating LLMs with laboratory automation. By interpreting technical documentation, Coscientist enhances its performance in automating experiments. This capability was extended to a more diverse robotic ecosystem, such as the Emerald Cloud Lab (ECL), demonstrating Coscientist’s adaptability and potential for broad scientific application.
According to the research paper, Coscientist’s real-world testing involved conducting experiments using the Opentrons OT-2, a liquid handler with a well-documented Python API. Through simple natural language prompts, the system could execute accurate protocols and integrate multiple hardware tools, showcasing its practical applicability in a laboratory setting.
“Beyond the chemical synthesis tasks demonstrated by their system, Gomes and his team have successfully synthesized a sort of hyper-efficient lab partner. They put all the pieces together and the end result is far more than the sum of its parts — it can be used for genuinely useful scientific purposes,” National Science Foundation Chemistry Division Director David Berkowitz said in a press release.
Coscientist’s reasoning capabilities were evident in its ability to plan and execute complex chemical experiments, such as catalytic cross-coupling experiments. It successfully designed high-level working protocols using Python, demonstrating its potential in advanced scientific research. This adaptability was further shown in its performance across various organic transformations, indicating its usefulness in exploring multiple chemical reactions.
The team is aware that Coscientist’s development raises important considerations regarding the ethical and responsible use of AI in scientific research. While it offers significant potential for advancing research, concerns about safety and the possibility of misuse exist. Addressing these concerns is crucial to harness the full potential of AI systems like Coscientist in scientific discovery while mitigating risks.
“I believe the positive things that AI-enabled science can do far outweigh the negatives,” said Gomes. “But we have a responsibility to acknowledge what could go wrong and provide solutions and fail-safes.”
“By ensuring ethical and responsible use of these powerful tools, we can continue to explore the vast potential of large language models in advancing scientific research while mitigating the risks associated with their misuse,” the authors concluded in the research paper.