2026-08-14
Le Song, Eran Segal, and Eric Xing outline a vision for an 'AI-driven digital organism': modular multiscale foundation models meant to simulate biology from molecules to individuals.
Biology underpins medicine, pharmacology, public health, and longevity research, but wet-lab work on living systems is slow, expensive, and often too risky for open-ended trial and error. In this Nature Medicine Perspective, Le Song (GenBio AI), Eran Segal (Weizmann Institute of Science), and Eric Xing (MBZUAI / GenBio AI) lay out a vision for using AI to model and simulate biology as an alternative or complement to wet-lab experimentation.
The paper's central proposal is the AI-driven digital organism (AIDO): an integrated system of multiscale foundation models, built in a modular and connectable way that mirrors biology's own layered, interconnected structure, from molecules up through cells to whole individuals. The abstract doesn't specify implementation details, such as which architectures cover which scale or how the modules would be aligned; those specifics would require the full text, which sits behind Nature's paywall.
This is a Perspective piece, not an empirical study, so it reports no experiments, benchmarks, or numbers. What it offers is an architectural vision for what such a system should look like, not results from a working prototype.
If something like AIDO gets built, the pitch is that prediction, simulation, and programming of biology could move from physical experiments to a computational platform that's safer, cheaper, and higher-throughput, in turn guiding which wet-lab experiments are worth running. It extends the broader virtual-cell line of work: generative models that simulate how a cell or organism responds to interventions rather than predicting a single fixed task. With authors spanning AI research (Xing and Song at GenBio AI) and genomics and precision medicine (Segal), Perspective pieces like this tend to shape where funding and follow-up research effort go next.
This is where honesty matters most: Nature Medicine paywalls the full text, so this summary is built entirely from the public abstract. Method details, whether any working prototype exists, and what evaluation criteria the authors propose are all unverifiable from what's public. A common risk with vision papers is that integrating multiscale foundation models can stay a diagram-level claim; the abstract gives no evidence about whether molecular, cellular, and organism-level models have actually been aligned into a working system. Readers who need to judge feasibility will need institutional or subscription access to the full text.