AI co-scientists are revolutionizing how research is done

Biochemist Anna Pertl submitted her query and departed for the evening, unaware that an artificial intelligence system, named Co-Scientist, was about to embark on a scientific journey. Unlike a typical chatbot, Co-Scientist employs multiple autonomous AI agents to scour the scientific literature, evaluate competing hypotheses, refine and critique ideas, and rigorously test concepts against published evidence.

The process is time-consuming, often requiring the equivalent of a long-distance triathlon for the researcher, but it yields innovative scientific insights.

During a Tuesday in July, Pertl tasked Co-Scientist with devising non-obvious, practical methods to harness the biology of molecular condensates to inhibit MYC, a protein that drives cancer cell proliferation. Co-Scientist, developed by Google, took some time to understand Pertl's query, initially assuming that the condensates were the drug target rather than the biological mechanism to be targeted. However, after some back-and-forth, the AI tool generated 108 possible strategies, of which only one proved viable.

The strategy proposed by Co-Scientist was to glue together the condensates using click chemistry, a molecular linking technology. This approach stiffens the condensates, hindering MYC's ability to turn on genes. This conceptually compelling strategy had never been considered before, as scientists were already aware of condensates' ability to transition from a liquid to a more rigid state. Researchers at companies like Dewpoint Therapeutics are now exploring this approach to target cancer cells.

AI-assisted brainstorming is becoming increasingly common, with organizations such as Frontier AI labs, Anthropic, OpenAI, FutureHouse, and Phylo developing systems that can tackle tasks traditionally reserved for human scientists. This collaboration between humans and AI has the potential to revolutionize research, freeing researchers to focus on high-level questions and decisions while AI takes care of the routine work.

However, this shift in the research process also presents a paradox. As machines handle more of the discovery process, the human component - knowing which questions are worth asking and which lines of inquiry are most promising - becomes even more critical. This could lead to researchers needing deeper knowledge to evaluate AI-generated hypotheses, potentially making it harder to train the next generation of scientists.

Nonetheless, those with human skills that understand the full picture will likely be more valuable than ever.