AI Science Assistants Succeed in Drug-Repurposing Tasks
▼ Summary
– Nature published two papers on AI systems designed to help scientists develop and test hypotheses: Google’s Co-Scientist uses a “scientist in the loop” model, while FutureHouse’s system can evaluate biological data from specific experiments.
– Both systems focus on biological data and straightforward hypotheses, aiming to assist rather than replace scientists or the scientific process.
– Both are “agentic” systems that operate in the background by calling out to separate tools, with Microsoft taking a similar approach and OpenAI being an exception.
– The systems address the challenge of the overwhelming volume of scientific information, as the explosion of journals and papers makes it hard for researchers to stay current in their field.
– Finding relevant material across different fields is difficult, such as a signaling system in eye development also being involved in kidney research.
On Tuesday, two separate papers in Nature unveiled AI systems designed to assist scientists in forming and testing hypotheses. One, Google’s Co-Scientist, operates on what the team calls a “scientist in the loop” model, meaning researchers regularly apply their own judgment to guide the system’s work. The other, developed by the nonprofit FutureHouse, goes a step further by training its AI to evaluate biological data from specific types of experiments.
Google says its system could eventually apply to physics, but both groups focus exclusively on biology and present relatively straightforward hypotheses, such as “this drug will treat that condition.” These tools are not intended to replace scientists or the scientific method. Instead, they aim to handle what current AI does best: sifting through massive volumes of information that would overwhelm any human.
So what is this technology actually good for? Both systems are agentic, meaning they run in the background and call on separate tools to complete tasks. (Microsoft has adopted a similar approach with its science assistant; OpenAI appears to be an outlier, having simply tuned a large language model for biology.) While there are differences between the two, they target the same core problem: the overwhelming flood of scientific literature.
The ease of online publishing has caused an explosion in the number of journals and papers. It is now extremely difficult for any researcher to keep up with their own field, let alone find relevant work in other disciplines. For example, someone studying eye development might miss that a signaling pathway they rely on is also active in the kidney, simply because they cannot track discoveries across every domain. These AI assistants are designed to bridge that gap.
(Source: Ars Technica)
