Blog

The virtual cell is having its moment: a Cell paper by forty-plus authors proposed building an AI Virtual Cell with multi-scale foundation models. But a patient is not a bigger cell—from cell to human body lie multiple emergent transitions, where pathway redundancy and patient heterogeneity keep making 'works in vitro' end at 'fails in vivo'. This essay argues that the path from virtual cell to virtual patient is missing not a bigger model, but a map at another scale—the 'molecule → module → human' conversion layer—and that virtual patients require two things virtual cells cannot offer: the medical semantics of Action (will this person respond to this intervention?) and a re-test feedback loop (predict → intervene → re-test → correct). The essay closes with the economics: the Phase II valley of death, ~$0.9 billion per approved drug, and biomarker stratification halving costs—the key is not under the streetlight.

After more than two decades of data arms race, biomedicine received a cold verdict: data explosion, knowledge poverty. Knowledge has long lived in three forms—travelogues (literature reviews), natural history museums (databases), and maps (computable knowledge)—and biomedicine lacks precisely the third. This essay argues from the history of science that a discipline matures when its core knowledge turns from narrative into computable objects: astronomy has star catalogs, chemistry has the periodic table, biomedicine has yet to build its coordinate system. It is time to attack the quantification of knowledge: parameterize, structure, and close the loop—so that a worldview finally acquires a computable carrier.

Disease is not sudden; it is the result of capacity reserve progressively collapsing along five lines of defense. Starting from the first principle that 'life is a collection of capacities to adapt to the environment,' and drawing on landmark studies on Aging Hallmarks, epigenetic clocks, organ aging, immunosenescence and metabolic imbalance, this article derives a Five-Lines-of-Defense framework for healthy aging: phenotypic early warning, organ function, metabolic homeostasis, immune repair, and Aging Hallmarks. The essence of health management is to continuously monitor and repair capacity reserve before disease becomes clinically manifest.

On May 8, 2026, the China National Intellectual Property Administration granted the invention patent Method, System and Application for Generating Compound Intervention Schemes Based on Pre-trained Models (CN117766054B). This patent corresponds to our technical exploration around SEMO: how to put compounds, protein-protein interaction networks, individual omics data, and intervention scheme generation into a computable framework. I prefer to understand it as a methodological shift from checking whether an indicator is below the average to examining whether there exists an intervenable gap within the individual network.

Over the past decade, AI in medicine has followed a clear migration path: Medical AI solves the problem of understanding medical facts, AI drug discovery solves the problem of finding intervention tools, and medical world models solve the problem of verifying intervention outcomes. The first two primarily change efficiency; the latter may change the medical paradigm itself. Medical AI is the past, AI drug discovery is the present, medical world models are the future.

What medical AI truly lacks may not be yet another larger model, but an ImageNet-like infrastructure — a foundational data and evaluation system that systematically records current biological states, intervention actions, and subsequent state changes. The next decade of medical AI does not lack large models. What is truly missing is a shared infrastructure for biological state transitions. Whoever defines state, action, and transition may define the underlying coordinate system of next-generation AI in medicine.

Medical AI cannot rely solely on the external guardrails of harness engineering. A true medical world model must enter the internal architecture of state-action-transition-feedback, making biological state changes representable, deducible, auditable, and correctable. Harness engineering controls the AI system from outside. Steerable world modeling structures biomedical reasoning from inside.

Prediction models answer what will happen; world models answer what will happen if I take this action. Medical AI must learn not just to predict, but to reason about the consequences of actions. The core structure of a world model consists of five elements: State, Action, Transition, Objective, and Feedback. The future competition in medical AI is not just about bigger models, more data, or better predictions, but about who can represent living systems as explainable, deducible, feedback-driven, and calibratable world models.

If a person stood before you and a world model stood beside them, would you trust it with your health? SEWO (Steerable Medicine World Model) argues that the ultimate mission of a world model is not to be a more accurate oracle, but to be a reliable steed — the doctor gives direction, and the model navigates uncertain terrain on its own. Prediction is the means; reliable state transition is the goal.

The same inflammation concept, measured via serum protein, shows weak correlation with brain health. Switch to a DNA methylation surrogate, and the correlation jumps 6.4x. The concept didn't change — the 'shutter speed' did. This is precisely the core insight behind EvoSika's methodology.

The mission of AI4S should not be using AI to accelerate existing science, but using AI to reinvent how science organizes knowledge. The content of a concept depends on the context of inquiry — we need an operating system for concepts.