A 22-Author Team Argues AI Must Simulate Biology, Not Just Read It
A new paper argues that AI in biomedicine must move from pattern recognition to forward simulation — and lays out the full blueprint for doing it
- Core proposal: A 22-author perspective paper argues biomedical AI must shift from static pattern recognition to forward simulation of biological futures using biomedical world models.
- How it works: Models encode biological states into a latent space, then predict how those states evolve under interventions (drugs, gene edits, surgery) before any real-world action is taken.
- Four key applications: Virtual cells, virtual organoids, virtual patients, and surgical simulation — spanning molecular to clinical scales.
- Major bottlenecks: Intervention data is sparse and unstructured; cross-scale integration from cell to organism remains unsolved; privacy rules block data centralization.
- No code or model released: This is a conceptual framework paper currently under review, not a system with benchmarks or checkpoints.
- Industry context: CZI, Google DeepMind, and others are already investing heavily in virtual cell platforms; this paper provides a unifying paradigm across those efforts.
Most AI systems in biomedicine today are, at their core, very sophisticated lookup tables. Feed them an image, a gene expression profile, or a clinical note, and they return a label or a score. Although foundation models and large language models have accelerated biomedical data interpretation, most current systems remain focused on static pattern recognition rather than prospective simulation of biological futures. A new perspective paper from a 22-author team argues that this is the wrong paradigm entirely , and proposes a concrete alternative.
The authors propose biomedical world models as a paradigm for AI-driven discovery. These models learn latent representations of molecular, cellular, tissue and clinical states, together with intervention-conditioned dynamics that allow future trajectories to be simulated before actions are taken. In plain terms: instead of asking "what does this data look like?", these models ask "if I apply this drug, perform this surgery, or knock out this gene , what happens next?"
The gap between knowing and predicting
The distinction matters enormously in practice. Biological systems are inherently dynamic, evolving through multiscale processes that respond to genetic perturbations, environmental changes and therapeutic interventions. Understanding such systems therefore requires models that can simulate state evolution rather than merely interpret observations. A model that classifies a tumor as "high-grade" is useful. A model that simulates how that tumor evolves under three different treatment protocols , and ranks them by predicted outcome , is transformative.
Drug discovery is increasingly adopting computational and data-driven methods to address persistent challenges, including high attrition rates, lengthy development timelines, and escalating costs. Pfizer recently found that model-informed drug development saved on average 10 months of cycle time and $5 million per programme. The promise of world models is to push those gains much further by enabling true forward simulation before any wet-lab resources are committed.
How a biomedical world model actually works
Formally, a world model is an internal, generative or predictive model that learns latent representations of a system and simulates how its states evolve over time. Given historical observations, the model infers a latent state that summarizes the current state of the system. Conditioned on previous actions or interventions and contextual variables, it then predicts future state transitions, capturing how the system may evolve under a specific action or intervention.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
