Harvard Team Publishes Biomedical World Models Framework in Cell

Marinka Zitnik's team at Harvard Medical School and the Kempner Institute published "World models for biomedicine" in Cell's "AI in Biology" special issue, co-led by ayushnoori and njwfish. The paper systematically brings the AI concept of world models—representing system state and internally simulating its evolution before acting—into biomedicine across scales from molecules to patients. Its core claim: biology needs models that predict what happens after an intervention, not models that merely describe the world.

Confirmed

• The paper appears in Cell's "AI in Biology" special issue, authored by researchers from Harvard Medical School, the Kempner Institute and collaborators, led by marinkazitnik's group and co-led by ayushnoori and njwfish
• The intellectual lineage runs from Kenneth Craik's idea that organisms test actions in small-scale internal models of reality, through robot planning and reinforcement learning, to world-model work such as Ha et al.
• The framework covers applications including virtual drug trials, inverse intervention design, and sequential planning
• Marinka Zitnik proposes a unified test: does the model represent system state, accept specified actions, predict resulting states, and continue reasoning from predictions? On this basis, foundation models can serve as encoders but are not themselves world models; digital twins that only mirror a single system's current state do not qualify; emerging efforts like virtual cells must likewise be judged by the four criteria
• The paper details the data types, training approaches, and evaluation methods needed to build biomedical world models, positioned as a framework reference for researchers at the AI–biology intersection

Why it matters

• The authors identify three key challenges: biological states are typically only partially observable; destructive measurements limit repeated measurement of the same system; and intervention data is scarce, making causal learning of intervention–outcome relationships difficult
• The four criteria provide a unified framework for evaluating foundation models, digital twins, and virtual cells, helping clarify debates over what counts as a world model
• If world models can simulate intervention outcomes before experiments, they could substantially reduce the cost and risk of biomedical experiments and accelerate translation from basic research to clinical application

2026-09-18 ~ 2026-09-18 · 8 related posts

• Cell Focus: Biology Needs World Models That Predict Interventions, Not Just Descriptions — rishabh16_ · 2026-09-18
• [source] Cell paper lays out how world models could transform biomedicine, from molecules to patients — marinkazitnik · 2026-09-18
• [source] Biomedical world models: a framework for virtual drug trials, intervention design and planning — marinkazitnik · 2026-09-18
• What makes a world model? A four-part test that foundation models, digital twins fail — marinkazitnik · 2026-09-18
• Three Open Challenges for Biomedical World Models, Per Marinka Zitnik — marinkazitnik · 2026-09-18
• [source] New Cell Paper: World Models to Simulate Biomedical Interventions Before Running Them — marinkazitnik · 2026-09-18
• Harvard team's Cell paper lays out three data challenges for biomedical world models — marinkazitnik · 2026-09-18

1 near-duplicate retellings: marinkazitnik