AI agents enter the ‘wet lab’
https://www.nature.com/articles/d41586-026-03270-1?utm_source=x&utm_medium=social&utm_campaign=nature&linkId=64326852
AI Agents Enter the Wet Lab: The Rise of Autonomous Scientific Discovery
From computational AI to experimental AI: AI is moving beyond literature reviews, data analysis, and molecular design toward physically conducting laboratory experiments with minimal human intervention.
Introducing AutoDNA: Wu et al. describe an AI-native autonomous laboratory in Nature, where multiple AI agents collaborate to plan, execute, and analyze molecular biology experiments involving DNA.
Multi-agent scientific collaboration: Specialized agents handle literature searches, reagent verification, protocol development, hypothesis generation, code generation, and instrument operation, coordinated by a central planning agent.
Integration with laboratory robotics: AutoDNA connects more than 20 types of laboratory instruments, including liquid handlers, thermal cyclers, and DNA sequencing systems, translating natural-language instructions into executable experimental workflows.
Demonstrating experimental capabilities: The system optimized an enzyme-mediated DNA synthesis reaction, achieving a 97.8% yield, encoded digital information into 78 DNA strands and successfully recovered it through sequencing.
AI-driven molecular diagnostics: AutoDNA also identified a cancer-associated KRAS mutation from a colorectal cancer sample by designing and synthesizing primers and experimentally confirming the mutation.
Beyond predefined protocols: One particularly interesting capability was the agents independently recognizing that reaction tubes needed to be capped during incubation to prevent evaporation—demonstrating an ability to account for practical experimental details not explicitly specified by the user.
The emergence of closed-loop science: Autonomous laboratories could bridge the growing gap between rapid AI-driven hypothesis generation and the slower process of experimental validation, enabling iterative cycles of experimental design, execution, measurement, and optimization.
Challenges and limitations: Current systems still require human intervention for experimental setup, reagent replenishment, troubleshooting, and waste management. Hardware errors, instrument compatibility, reproducibility, and the limited range of automatable experiments remain important barriers.
The future of autonomous science: Digital twins, simulation-based validation, standardized instrument interfaces, operational safety, and biosecurity screening will be essential for reliable deployment. Ultimately, integrating computational AI agents with robotic experimental platforms could create self-improving scientific laboratories that accelerate discovery across molecular biology, biotechnology, drug discovery, and biomedical research.
The bigger picture: We are moving from AI that helps scientists think about experiments to AI that can actually perform them. The convergence of large language models, multi-agent systems, laboratory robotics, and experimental feedback could fundamentally transform how scientific knowledge is generated.