AI Scientific Assistants Rise: Google’s Co‑Scientist and FutureHouse’s Robin

Co‑Scientist (Google DeepMind)

Co‑Scientist is built on Gemini 2.0 and uses a multi‑agent architecture consisting of a generation agent, a reflection agent, a ranking agent, an evolution agent, and a meta‑review agent. The system integrates web‑search and domain‑specific databases and permits scientists to intervene, provide feedback, or supply hypotheses.

Design criteria are reasonableness, novelty, testability, and safety. The reflection agent can query external search tools, which the authors report reduces hallucinated hypotheses that lack logical consistency.

In an acute myeloid leukemia (AML) drug‑repositioning task, the system screened 2,300 approved drugs. After expert review, candidates including Binimetinib, Pacritinib, and Cerivastatin were selected for experimental testing. Across multiple AML cell lines these compounds showed very low IC50 values and higher selectivity versus non‑AML lines. The system also generated previously unexplored drug combinations that produced strong synergistic effects in the MOLM‑13 cell line.

In a blind expert evaluation of eleven open biomedical questions, Co‑Scientist received the highest scores for novelty, impact, and overall preference.

Paper: https://www.nature.com/articles/s41586-026-10644-y

Robin (FutureHouse)

Robin’s workflow comprises three specialized agents: “Crow” for literature search, “Falcon” for literature review, and “Finch” for experimental data analysis. Finch can ingest raw wet‑lab data, generate and execute Jupyter‑notebook code, perform gating strategies, conduct differential‑expression analysis, and automatically produce figures and statistical summaries.

Targeting dry age‑related macular degeneration (dAMD), Robin processed 551 papers in ~30 minutes, extracting ten disease mechanisms and proposing thirty candidate drugs focused on retinal pigment epithelium (RPE) phagocytosis. Falcon evaluated each candidate, and a large‑language‑model judge performed pairwise comparisons to rank hypotheses, analogous to a tournament system.

In the first experimental round, inhibitors suggested by Robin enhanced phagocytic activity. Robin then initiated RNA‑seq on treated cells; Finch’s autonomous analysis yielded new findings that drove subsequent hypothesis updates. The entire pipeline, covering roughly 825 reference papers, required ~30 minutes of compute time, compared with an estimated >800 hours of manual expert effort. All statistical plots in the publication were generated automatically by Robin.

Paper: https://www.nature.com/articles/s41586-026-10652-y

Position of AI in Scientific Research

Replacing the literature‑search agent Crow with OpenAI’s o4‑mini increased hallucinated citation rates from 0 % to 45 %, demonstrating the importance of tools specifically designed for scientific literature interfacing.

Both systems stop short of fully automated science; their successes are limited to tractable stages such as in‑vitro screening and do not involve de‑novo molecule design. The authors note that many drug candidates still fail in animal or clinical trials despite promising cell‑culture results, and that AI currently cannot resolve underlying mechanisms of efficacy or explain gene‑expression changes.

Code example

来源：ScienceAI
本文
约2000字
，建议阅读
5
分钟
多智能体架构、实验室闭环——AI正在从执行者变为合作者。