Yongsheng Intelligent Launches ProtoPilot and BioLab Bench to Advance AI for Bio with Closed-Loop Experimentation

ProtoPilot, jointly released by Yongsheng Intelligence and Shanghai Artificial Intelligence Laboratory, is the world's first self-evolving multi-agent system in a real laboratory setting. It completes the full chain from natural language experimental requirements, Protocol generation, SOP synthesis, automation code conversion, to device execution, experimental feedback, and plan optimization. In experiments such as bacterial culture, colony PCR, plasmid construction, site-directed mutagenesis, and DNA assembly, ProtoPilot has executed generated workflows on real devices and can modify plans based on failure results and re-run experiments.

The key breakthrough lies in solving the long-standing pain point of AI being able to think but not do. Yang Meng points out that AI can generate seemingly reasonable experimental plans, but it is difficult to convert them into machine-understandable and executable actions, let alone summarize causes and adjust after real wet lab failures. ProtoPilot incorporates physical constraints (such as samples, consumables, volumes, temperatures, well positions, and device interfaces) into the system, giving AI "hands" and "eyes"—devices can be driven by code, experimental results can flow back, failures and expert judgments are continuously accumulated, and the system learns and evolves through iteration.

Simultaneously released with ProtoPilot, BioLab Bench is the first full-process Agent evaluation system in the life sciences field, from user requirements to device execution. It measures whether an agent can convert experimental intent into plans, SOPs, and machine-executable code, adapt to different brands of automation equipment, and ultimately be validated through real experiments. Yang Meng emphasizes that BioLab Bench assesses not just whether AI understands experiments, but whether AI can complete experiments.

Although tech giants like Google DeepMind, OpenAI, and Anthropic have successively entered the life sciences field, Yang Meng believes they are still in the dry lab stage and have not truly entered the laboratory. OpenAI and Anthropic are more focused on showcasing visions to the capital market based on their own model capabilities; DeepMind has the deepest foundation, with AlphaFold winning a Nobel Prize. However, if model giants build their own labs, they may adopt a top-down AI-native design path.

Yongsheng Intelligence has chosen a different path: bottom-up, starting from specific scenarios. Yang Meng states that the company currently focuses on areas like targeted sequencing library preparation, which require extensive trial and error, lengthy experimental processes, and high demands for standardization and reproducibility. The team uses foundation models for Harness Engineering, converting expert experience into sustainably accumulated knowledge solidified in the AI system. As more employees and partners use it, the "human-in-the-loop" knowledge accumulation mechanism gradually improves.

Yang Meng believes that both paths ultimately aim for the same goal—achieving a dry-wet experimental closed loop. However, influenced by capital scale, narrative logic, and business models, model companies tend to be top-down, while Yongsheng Intelligence is bottom-up. He admits that currently no one has truly run through the entire process of life science research; end-to-end remains an ideal state, but the future will certainly move in that direction.

Yang Meng points out that large models encounter similar issues when entering the real physical world. Laboratories need to interact directly with the physical world, maintaining AI's intelligence while ensuring safe operation within clear boundaries and reducing hallucinations—this is inherently complex engineering. Manual experimental workflows cannot be directly transferred to machines: SOPs that humans can execute smoothly may not be suitable for automation equipment; liquid handlers have their own capability limits in pipetting accuracy and operation methods. Transitioning from manual to automation is not a simple translation but requires iterative optimization.

Manipulating liquids, cells, and microorganisms with machines involves complex issues such as sensors and control logic. Only through long-term engineering and product accumulation, exposure to enough experimental protocols, liquids, consumables, microorganisms, and cells, can we push laboratories from automation to intelligence and autonomy. Yang Meng emphasizes that every device operation is backed by profound logical principles and practical experience—this is true industry know-how.

Yongsheng Intelligence's ultimate goal is to deliver experimental results, not just sell equipment. Therefore, the company considers not the productization of a single module, but how to make the entire life science workflow sufficiently intelligent and call upon corresponding automation equipment at each step. Yang Meng believes that what limits AI's development in life sciences is often not model capability, but real-world hardware operations—physical devices have errors, and the transition between device processes is extremely complex.

With the addition of AI, the laboratory composition evolves into "human + device + AI." Yang Meng believes that a better path is to set sufficient rules and constraints for AI in each module, and through positive human-machine interaction, let AI assist humans in reaching blind spots they had not considered before. In the short term, fixed process operations can be handed over to robots, but every decision and feedback step still requires human oversight, and should be completed collaboratively by humans and AI.

Yang Meng points out that individual human knowledge and experience are limited. In the past, when anomalous data appeared in experiments, it was often discarded as "systematic error." But in the AI era, we hope for humans and AI to jointly analyze and precisely determine whether anomalous data is instrument noise or a potential new scientific phenomenon. This collaborative exploration model is the most fundamental difference between the AI-era laboratory and the past.

Regarding the trust issue caused by AI's "black box," Yang Meng proposes solutions from two dimensions: top-down and bottom-up. Top-down: define strict Rubrics (acceptance and evaluation rules) for AI, converting accumulated human judgment rules into explicit constraints. Bottom-up: when encountering out-of-distribution new data, introduce a "third perspective"—use a base large model not contaminated by biased data to make cross-judgments based on first principles of physics and science.