Biology is the foundation of medicine, pharmacy, public health, and longevity research. It is also, by nature, extraordinarily difficult to work with. Living systems are complex, experiments are expensive, and interventions carry real risk. A perspective published in Nature Medicine proposes a way to change that equation: build an AI-driven digital organism, a computational system capable of modeling and simulating biology at every scale, from individual molecules to whole human beings.
This is not a product announcement. It is a research vision, and understanding what it proposes, and why it matters, requires stepping back from the headline.
Biology Is Too Complex to Tamper With Freely
The core problem the authors identify is straightforward. In the physical world, biological systems resist easy manipulation. Running experiments on living cells, tissues, or organisms takes time, money, and carries ethical and safety constraints. This limits how quickly researchers can test hypotheses, explore drug candidates, or understand disease mechanisms.
The proposed solution is a concept called an AI-driven digital organism, or AIDO. The idea is to construct a system of integrated, multiscale foundation models, AI architectures trained on biological data at different levels of organization, and connect them in a modular, holistic way that reflects how biology actually works. Molecules interact with cells. Cells form tissues. Tissues make up individuals. An AIDO would model all of these layers and, critically, model how they connect.
The vision is a platform that is safe, affordable, and capable of high-throughput simulation. Researchers could run thousands of virtual experiments that would be impractical or impossible in a wet lab, then use those results to guide the physical experiments that actually get run.
Foundation Models, Applied to Life Itself
The technical backbone of this vision draws on a class of AI architecture that has already transformed natural language processing and computer vision: foundation models. These are large models pre-trained on vast datasets, capable of being adapted to many downstream tasks. The authors point to a range of existing work that applies this approach to biological data.
Several models already exist for specific biological domains. Tools like the Nucleotide Transformer and HyenaDNA address genomic sequences. Evo extends sequence modeling from the molecular to the genome scale. On the protein side, AlphaFold and AlphaFold 3, developed by researchers at DeepMind and published in Nature, have demonstrated highly accurate prediction of protein structures and biomolecular interactions. Models like xTrimoPGLM, a 100-billion-parameter transformer, have been developed specifically to decode the language of proteins. At the cellular level, tools like scGPT and scLong target single-cell transcriptomics, with scLong described as a billion-parameter model designed to capture long-range gene context.
What the AIDO vision proposes is not to build one more specialized model, but to integrate these layers into a coherent, connected system. The authors describe this as modular and connectable, meaning individual components can be developed and improved independently, but they are designed to communicate with each other. This reflects a key insight: biology does not operate in isolated layers, and a useful simulation of biology cannot either.
The research group behind this perspective is affiliated with GenBio AI, based in Palo Alto, and includes researchers Le Song, Eran Segal, and Eric Xing.
What It Would Mean to Simulate a Living System
Here is what most coverage of AI in biology misses. The ambition here is not to automate existing experiments. It is to create an entirely new kind of scientific instrument, one that operates in silico rather than in a lab, and that can be queried, programmed, and explored in ways that physical biology does not allow.
The authors describe the AIDO as a platform for predicting, simulating, and programming biology. That last word is significant. Programming implies not just understanding a system but being able to specify desired behaviors and work backward to the interventions that would produce them. In drug discovery, that could mean designing molecules with target properties. In medicine, it could mean modeling how a specific patient’s biology might respond to a treatment before that treatment is administered.
The authors are careful to frame this as complementary to physical experimentation, not a replacement for it. The vision is that an AIDO would trigger better-guided wet-lab work and better-informed reasoning from first principles. Researchers would use the digital system to narrow the space of possibilities, then test the most promising candidates in the real world.
This matters beyond the laboratory. If biological simulation becomes reliable and accessible, it changes the economics of drug development, the speed of public health response, and the depth of understanding that researchers can bring to complex diseases.
In Short
An AI-driven digital organism is a proposed system of interconnected AI models, each trained on a different layer of biological data, designed to simulate life from molecules to whole individuals. The goal is not to replace physical biology but to create a computational platform where hypotheses can be tested safely, cheaply, and at scale, making physical experiments more targeted and more likely to succeed. The building blocks, foundation models for genomics, proteins, and single cells, already exist. The challenge now is integration.
Based on reporting from Nature: Machine Learning.