TLDR
A computer partner generated promising scientific ideas by searching information, criticizing its own suggestions, and improving them repeatedly. Some ideas worked in laboratory tests, but human scientists still need to check them carefully before real-world use.
Summary
This content was automatically synthesized by Credo's AI models directly from the original source text.
AI summaries can make mistakes — double-check important details against the original source.
1 Study Aim
The study aimed to develop and test Co-Scientist, a multi-agent artificial intelligence (AI) system (computer agents working together), for generating novel, testable scientific hypotheses. The authors asked whether increasing its reasoning time and computational effort would improve hypothesis quality and whether expert scientists could use it across biomedical problems. The system was designed to help scientists find and test useful new ideas faster.
2 Study Design
Co-Scientist used Gemini 2.0 models, an asynchronous task framework, persistent memory, literature searches, and specialized agents for generation, reflection, ranking, evolution, and review. An Elo-based tournament (a repeated contest ranking ideas against each other) compared hypotheses as the system used test-time computation (extra processing during an answer). Evaluations covered 203 research goals, 15 expert-curated biomedical goals, and 11 blinded expert assessments. Laboratory tests examined five cell lines in biological triplicates. The team studied acute myeloid leukemia (AML), liver fibrosis, and antimicrobial resistance. The researchers combined computer comparisons, expert judgments, literature checks, and preliminary laboratory experiments.
3 Findings
The study reports that later hypotheses improved across 203 goals as computation increased. Co-Scientist eventually outperformed comparison models and experts in automated rankings, while experts preferred its outputs in a small blinded assessment. In AML, laboratory tests supported activity from several repurposed drugs, including Binimetinib, Pacritinib, Cerivastatin, and the novel candidate KIRA6. Several proposed drug combinations were synergistic (stronger together) in some leukemia cell lines. The system also identified epigenetic targets (gene-control changes) for liver fibrosis, with two drugs reducing fibrosis in liver organoids (small lab-grown tissues). It proposed a mechanism for bacterial gene transfer that matched independent experimental work. The authors stress that these results are preliminary, limited by literature quality, model errors, small evaluations, and laboratory-only testing. The findings suggest faster idea generation, not a replacement for expert review, rigorous experiments, or clinical trials.