Biology in the Agentic AI Era
Human society is undergoing a silent yet rapid industrial revolution driven by large language models. Unlike earlier transformations driven by mechanization, electrification, or computation, this one is fundamentally about the democratization of human cognition. Agents are changing how artificial intelligence interacts with the world, and in doing so they are beginning to expand the boundaries of what we might call the form of life.
The life sciences are being profoundly reshaped within this broader transformation. AlphaFold showed that AI could generalize from known protein structures and thus achieve accurate structure prediction; RFdiffusion pushed further, demonstrating that generative models can design proteins de novo. In drug discovery, systems such as Isomorphic Labs' AI-native drug design engine and Insilico Medicine's AI-designed rentosertib Phase IIa clinical results suggest that AI is moving into the full pipeline from structure prediction to molecular design and therapeutic pipeline generation.
Yet the underlying force of this revolution is the exponential growth of data scale and data quality. Modern language models are trained on corpora at the scale of trillions of tokens; through agents, MCP servers, and self-improving loop engineering, LLMs are becoming more than passive systems that fit data distributions. They are beginning to actively construct their own operating environments and continuously extend their capability boundaries in an AI-native way.
The field of structural biology presents a sharp contrast. Although AlphaFold achieved remarkable generalization from the existing universe of solved structures, the true heterogeneity of proteins is not captured by these predictive models. Proteins undergo continuous conformational change, exist as multi-state ensembles, and transition dynamically in response to their environments. These essential functional properties remain difficult to characterize systematically, making it hard to link structure, mechanism, and practical experimental workflows.
Challenges of Scalable Structural Biology
A more fundamental question is now becoming visible: how can we systematically expand the data and cognitive infrastructure of structural biology?
In this context, cryo-EM is the only experimental technology currently capable of directly observing macromolecular structures and conformational changes at near-atomic resolution. But its extraordinary physical capability stands in sharp contrast to the fragmentation of its full analysis workflow.
From sample preparation and microscopy to motion correction, CTF estimation, particle picking, 2D classification, 3D reconstruction, and refinement, each stage depends on large amounts of experience-driven hyperparameter selection and manual tuning.
This process is not a computational pipeline in the strict sense. It is closer to a scientific craft system that depends on expert intuition and accumulated experience. Different protein systems, laboratories, and even individual researchers often develop highly heterogeneous bodies of tacit knowledge. This makes structural determination difficult to reproduce and scale.
Project Ariadne
Project Ariadne is our first attempt to build a new kind of AI agent system that can understand and participate in structural biology work.
The name Ariadne comes from Greek mythology. She gave the hero Theseus a thread that allowed him to find his way through a complex and dangerous labyrinth and return safely. Current structural biology is also a vast labyrinth, filled with branching paths, uncertainty, and decisions that depend heavily on experience. Ariadne does not provide the answer itself. It provides a way to move through complexity while preserving the ability to trace, revise, and recover the path.
In this system, Ariadne is not simply using existing tools. Its goal is to gradually transform the experience and judgment scattered across different laboratories into a unified knowledge base, and from that foundation build a decision-making capability that can continuously evolve.
By learning from many historical experimental workflows and outcomes, the system attempts to move the practical knowledge of how structural biology is done from the memories of individual experts into runnable agents.
In this process, an agent is no longer merely an interface for executing commands. It becomes an action unit capable of judgment and collaboration in complex experimental environments. A database is no longer merely a place to store data. It becomes a knowledge substrate that carries the collective experience of structural biology. And a scientific workflow gradually becomes a dynamic decision space that can be learned and optimized.
Building Together
Ariadne is not a finished system, but an early form of scientific infrastructure still being built with the community. We hope to work with structural biology laboratories, computational biology teams, and industry partners that can bring real experimental workflows, data feedback loops, and validation in practical discovery settings. If you are interested in advancing this direction with us, please contact us through the Ariadne survey or email us directly at contact@cellverse.tech.
About Us
Cellverse is a Shanghai-based team with deep roots in AI. We are building an AI-native company for structural biology, with the belief that the next generation of scientific infrastructure will be shaped by models, agents, and data systems that learn directly from real experimental work.
Over the past two years, we have consistently pushed the boundary of AI x structural biology through cryoGEM (NeurIPS 2024), DRACO (NeurIPS 2024), CryoACE (ICML 2026, dynamic atomic model building), and our latest work, cryoUNI (structural landscape analysis). Ariadne brings this direction into an agentic workflow system for real structural biology practice.