In the quest for scientific breakthroughs, researchers blend creativity and expertise with literary insight to develop innovative research paths and guide explorations. However, navigating the rapid increase in scientific publications while integrating knowledge from diverse fields presents a challenge. Addressing this issue is essential, as seen in modern discoveries from interdisciplinary efforts, such as the 2020 Nobel Prize in Chemistry awarded to Emmanuelle Charpentier and Jennifer Doudna for their CRISPR innovation, which integrated knowledge across microbiology, genetics, and molecular biology. To tackle the unmet needs in scientific discovery and leveraging advancements in AI, we developed an AI co-scientist system. This multi-agent AI, built on Gemini 2. 0, serves as a collaborative tool for researchers, enhancing the scientific method's reasoning process. The AI co-scientist goes beyond standard literature reviews; it aims to generate original knowledge and formulate unique research hypotheses tailored to specific objectives. The AI co-scientist employs test-time compute scaling to refine its output iteratively.

Key processes include self-play-based scientific debates for hypothesis generation, ranking tournaments for hypothesis evaluation, and an "evolution" strategy for quality enhancement. To ensure effectiveness, we utilized the Elo auto-evaluation metric from these tournaments, discovering that higher Elo ratings correlate with greater output accuracy. In drug repurposing efforts for acute myeloid leukemia (AML), our AI co-scientist suggested novel repurposing candidates, validated through computational biology, expert input, and laboratory tests. Additionally, we explored hypotheses related to antimicrobial resistance (AMR) by instructing the AI to investigate mechanisms of bacterial gene transfer. The AI independently proposed that capsid-forming phage-inducible chromosomal islands (cf-PICIs) interact with phage tails to broaden their host range, a finding subsequently validated through prior experiments. This work is a collaborative effort among various teams at Google Research, Google DeepMind, and Google Cloud AI, alongside partners from several institutions, including the Fleming Initiative and Imperial College London. We extend our gratitude to all contributors and experts providing essential feedback, as well as our internal teams whose support was integral to this endeavor.