The way scientific research is conceived is changing. AI has moved from being a passive tool to an active collaborator in the scientific process. A survey revealed that LLMs are fast emerging as an essential tool for doctoral students in their research.
“In the next couple of years, we will get to a place where we have an AI scientist. Then you can apply a much faster rate of scientific progress,” said venture capitalist Vinod Khosla in a recent interview while talking about AI’s potential in scientific discovery.
This vision is already taking shape. In a new study, Stanford researchers introduced the concept of ‘Virtual Lab’, an AI-human collaboration framework designed for interdisciplinary scientific research—in this case, designing and validating new nanobodies for SARS-CoV-2.
This system employs a team of large language model (LLM) agents, each specialised in various fields, to assist human researchers in addressing scientific problems.
Built in Python, it uses GPT-4o as its default LLM but supports easy replacement with other models—unlike other research frameworks like ChemCrow and Coscientist, which rely on fixed or predefined LLM agents.
In such a structure, the agents do more than just performing computations. They engage in discussions, propose solutions, and critically evaluate outcomes– making them more than mere tools in the scientific process.
Inside Virtual Lab
The multi-agent AI architecture described in this research is designed to mimic an interdisciplinary team of scientists. The workflow begins with defining roles for AI agents like principal investigators (PI), scientist agents, and a scientific critic. The research progresses through structured “team meetings” for broad discussions and “individual meetings” for specific tasks.
AI agents discuss, critique, and refine ideas collaboratively, using tools like AlphaFold, Rosetta, and machine learning models to design nanobodies for SARS-CoV-2 variants. The human researcher sets agendas and integrates results, streamlining complex, interdisciplinary research with the help of AI.
A “scientific critic” agent ensures rigorous evaluation, enhancing output robustness, while the “principal investigator” agent synthesises inputs, simulates leadership and overcomes challenges in coordinating knowledge across domains.
“The LLM agents in the Virtual Lab serve as an extremely helpful resource for producing research ideas and experimental workflows, and code for research projects. But, it is up to the human researcher to double-check the validity of the decisions made by the LLM agents,” Kyle Swanson, a researcher at Stanford and co-author of this study, told AIM, commenting on the importance of human-in-the-loop for such experiments, at least currently.
While LLM agents are valuable for generating research ideas, workflows, and code, researchers must verify their accuracy—as they are capable of hallucinations and errors.
Swanson explained that although using multiple agents and the scientific critic can assist in spotting errors, this method is not foolproof, as all agents might make the same mistake. He also noted that offering extra context, like relevant research papers, can help reduce hallucinations, but won’t completely eliminate them.
Future of Interdisciplinary Scientific Discovery
Google DeepMind’s recent essay too highlights AI’s growing role in science. AI speeds up knowledge-sharing with LLMs, helps generate and label data (like predicting protein functions), and boosts research, it said.
Tools like AlphaProof and AlphaGeometry 2 are solving tough problems in drug design and algorithm optimisation. In May 2024, DeepMind released AlphaFold 3, an open-sourced, game-changing protein folding model that predicts with 50% better accuracy. Demis Hassabis and John M Jumper, two Google DeepMind scientists, were awarded the 2024 Nobel Prize in Chemistry for AlphaFold2.
This Stanford study references Jumper’s AlphaFold 2 paper as an illustration of interdisciplinary research. It highlights how modern scientific research involves large, diverse teams, which shows the collaborative nature of complex scientific research.
Beyond Virtual Lab
Interestingly, the Virtual Lab architecture can be expanded to other fields in drug discovery and material science with broader agent expertise. It is designed to be easily adaptable to a wide range of scientific applications, and its core elements are domain-agnostic.
“We hope that the Virtual Lab will continue to improve as the underlying LLMs get better,” said Kyle. “We also hope to incorporate more tools to give it more capabilities,” he added, suggesting improvements such as powering internet searches or giving access to GitHub repositories, which could help it acquire new knowledge and utilise tools beyond its training.
Experiments like these democratise access to entry-level knowledge. Virtual lab designed 92 nanobody candidates across four rounds, which were then experimentally validated. This efficiency showcases how resource-limited settings can achieve impactful results using AI collaboration.
This post was originally published on here
