Google DeepMind and Edison Are Building the AI Scientist
Google DeepMind and Edison Scientific are on an ambitious mission to build the AI scientist. These platforms propose to automate the scientific method using reasoning systems that connect hypothesis generation, experimental design, and data interpretation in one platform. In drug discovery, where traditional development timelines can stretch beyond a decade, such systems promise to dramatically accelerate the pace of biomedical research.
The AlphaFold developer and the nonprofit home organization behind Edison, FutureHouse, originally introduced their respective systems, Co-Scientist and Robin, as bioRxiv preprints in early 2025. Those studies have now been published in Nature, marking another step toward a growing ecosystem of specialized AI agents for life science research.
Led by Demis Hassabis, PhD, CEO, and 2024 Nobel laureate in Chemistry, DeepMind is no stranger to expanding biomedicine. The team published a January Nature paper describing AlphaGenome, a unifying DNA sequence model for regulatory variant-effect prediction to support understanding of genome function and disease biology.
Additionally, DeepMind drug discovery spinout, Isomorphic Labs, recently made waves after securing a whopping $2.1 billion Series B led by Thrive Capital, signaling the industry’s growing investment in AI-driven therapeutics.
“I’ve always believed the No.1 application of AI should be to improve human health,” wrote Hassabis on LinkedIn when announcing Isomorphic’s blockbuster raise.
DeepMind’s newly published AI assistant, Co-Scientist, is a general-purpose multi-agent system built with Google’s Gemini and driven by natural language prompts. The platform demonstrated initial validation across three biomedical applications: drug repurposing for acute myeloid leukemia, novel target discovery for liver fibrosis, and explaining mechanisms of anti-microbial resistance.
Co-Scientist’s design scales test-time compute to iteratively reason, evolve, and improve the output as it gathers more knowledge. Researchers can also actively steer the system by refining generated ideas or providing feedback through the natural language chat.
Vivek Natarajan, research scientist at DeepMind, emphasizes that time is a valuable commodity when tackling disease. Co-Scientist aims to support humans scientists in reaching answers to their problems much faster than before, from “months and years to minutes and hours.”
“To realize this vision, we need to build in reliability, trustworthiness and ensure a collaborative human-AI interaction paradigm. We have done a lot of research on these aspects and we are continuing to improve,” Natarajan told GEN Edge.
Closing the loop
Edison is the commercial spinout of FutureHouse, an AI scientist non-profit backed by former Google CEO Eric Schmidt and co-founded by Sam Rodriques, PhD, former group leader at The Francis Crick Institute and Edison’s CEO. The team’s newly published platform, Robin, leverages both OpenAI o4-mini and Anthropic Claude 3.7 to aid biological discovery.
In research tasks, Robin proposed repurposing Ripasudil, an existing drug for treatment of glaucoma, to address dry age-related macular degeneration (dAMD) via a novel mechanism that enhanced retinal pigment epithelial cell phagocytosis. The platform also suggested a circadian clock modulator, KL001, as an unexpected treatment for dAMD, illustrating the ability to make new connections not found in existing literature. Both insights were experimentally validated in patient-derived retinal pigment epithelium (RPE) cells.
Since Robin’s May 2025 preprint release, Edison unveiled an updated AI scientist, Kosmos, last November. Kosmos can reason over 175 million full-text papers, clinical trials and patents, and operate interactively as a colleague that can sends updates mid-run. The system is reported to perform hundreds of research tasks in parallel to compress months of work into a single day.
Today, Edison announced a collaboration with Incyte to employ Kosmos across the global pharma’s discovery and development pipeline. The partnership will focus on enabling continuous learning from translational and clinical data, real-time synthesis of evidence, and predictive models of therapeutic performance.
Michaela Hinks, founding member of technical staff at Edison, says the main bottlenecks for AI scientist adoption are trust, validation, and the gap in end-to-end solutions.
“Most AI tools accelerate the cheaper and easier upstream work, but not the expensive and regulated downstream stages of scientific research,” Hinks told GEN Edge.
She also highlights Robin as the first demonstration of an agentic AI scientist generating a hypothesis that is tested and validated in patient-derived cells, not an immortalized cell line, supporting clinically actionable insights for patients in need.
Whether AI scientists will truly revolutionize discovery remains to be seen, but researchers are already beginning the experiment.
