Co-Scientist uses several specialized agents to generate, critique, rank and refine research ideas. Scientists set the research goal and can add constraints or feedback, while the system searches literature and uses external tools to improve its proposals.
How the two-day result was tested
Researchers asked Co-Scientist to explain how capsid-forming phage-inducible chromosomal islands, or cf-PICIs, occur across multiple bacterial species. These mobile genetic elements can carry virulence and antibiotic-resistance genes.
Given minimal background information, the system’s top-ranked hypothesis was that cf-PICIs interact with different bacteriophage tails to expand their host range. The paper says that answer, generated in two days, matched the mechanism found by a separate genomic and experimental study before that work completed peer review.
The setup matters. Researchers deliberately chose a problem their group had already solved but had not made public. That allowed them to test whether the AI could independently reconstruct the same explanation from existing open-access literature without exposing it to the unpublished result.
What the leukemia tests found
The team also used Co-Scientist to suggest existing drugs that might be repurposed for acute myeloid leukemia, or AML. In one experiment, experts selected five candidates for cell-line testing; three reduced cancer-cell viability. In a separate test without outside candidate inputs, experts chose three of the system’s top suggestions for laboratory evaluation.
One of those drugs, KIRA6, reduced viability in several AML cell lines. Its half-maximal inhibitory concentration, the dose needed to cut cell viability by half, was 10 nanomolar in KG-1a leukemia cells versus 180 nanomolar in a non-AML control. Other AML cell lines were much less sensitive, showing that the response varied by subtype.
What the study does not prove
These findings were produced in cells grown in the laboratory. The authors explicitly caution that such tests cannot capture patient differences, drug metabolism, off-target effects or the tumor environment, and do not guarantee success in animals or people.
The paper also lists broader limitations: Co-Scientist depends heavily on open-access literature, can inherit weak or contradictory research, and remains vulnerable to factual errors and hallucinations. Google’s own earlier description of the system likewise presented it as an expert-in-the-loop tool. The study supports using AI to accelerate the search for testable ideas, while leaving verification to scientists, experiments and peer review.
