Gemini for Science: AI experiments and tools for a new era of discovery

TL;DR

• Google launched "Gemini for Science," a suite of experimental tools designed to act as general agents for scientific discovery, moving beyond narrow specialized models.
• The initiative includes three primary prototypes: Hypothesis Generation (using Co-Scientist), Computational Discovery (using AlphaEvolve and ERA), and Literature Insights (using NotebookLM).
• New "Science Skills" integrate over 30 life science databases, enabling complex bioinformatics and genomic analyses to be performed in minutes rather than hours.
• Validation research for the ERA and Co-Scientist tools has been published in Nature, while enterprise partners like BASF and Klarna are already utilizing these systems in private previews.
• Google is collaborating with over 100 institutions and piloting agentic peer review tools like ScholarPeer at major conferences including ICML and NeurIPS.

Why It Matters

This release signifies a shift from AI as a passive search engine to an active participant in the scientific method, capable of generating and testing hypotheses autonomously. For practitioners, it offers a new category of "agentic" tools that can compress months of manual computational work into parallelized minutes. The industry relevance is high due to the immediate enterprise adoption in pharmaceutical and logistics sectors, suggesting that AI-driven R&D acceleration is transitioning from theoretical to operational reality.

Key Data

• 30: The number of major life science databases and tools (including UniProt, AlphaFold, AlphaGenome API, and InterPro) integrated into the "Science Skills" bundle.
• 100: The number of institutions collaborating with Google to validate the new systems, including Stanford University and Imperial College London.
• "Minutes rather than hours": The time reduction achieved in complex structural bioinformatics and genomic analysis workflows when using Science Skills.
• "Thousands": The number of code variations generated and scored in parallel by the Computational Discovery prototype to test novel modeling approaches.
• "Nature": The publication venue where the validation papers for the ERA and Co-Scientist research tools were published.

Technical Details

• Multi-Agent Idea Tournament: The Hypothesis Generation tool employs a multi-agent framework where AI agents generate, debate, and evaluate scientific hypotheses to simulate the scientific method, with outputs supported by clickable citations for rigor.
• Parallel Code Generation: The Computational Discovery engine uses AlphaEvolve and ERA to generate and score thousands of code variations in parallel, allowing for the rapid testing of modeling approaches in fields like solar forecasting and epidemiology.
• Database Integration: The Science Skills bundle programmatically links AI agents to external scientific data repositories such as the AlphaFold Database and UniProt, enabling automated retrieval and analysis of genomic and structural data.
• Agentic Peer Review: Experimental tools like the Paper Assistant Tool (PAT) and ScholarPeer are being piloted to perform agentic peer review, utilizing AI to assist in the validation and critique of scientific manuscripts at conferences like ICML and NeurIPS.

Industry Insight

• Shift to General Agents: The strategic emphasis on "general agents" rather than narrow models suggests that the next competitive advantage in AI will come from versatile systems that can adapt to diverse scientific domains without requiring domain-specific fine-tuning for every new task.
• Enterprise R&D Monetization: The active use of these tools by industry leaders like BASF and Klarna indicates that AI-driven operational efficiency in supply chain and ML optimization is a viable immediate revenue stream for AI providers, distinct from the longer-term academic benefits.
• Integration of Validation: By embedding Nobel laureates and PhD students in a "trusted tester community," Google is addressing the critical trust barrier in AI-generated science, positioning verification as a core product feature rather than an afterthought.

zations are currently using these tools in private preview?
A: Companies such as BASF and Klarna are using AlphaEvolve, while Daiichi Sankyo, Bayer Crop Science, and the U.S. National Labs are using Co-Scientist for research acceleration.