This journal exists to think in public about one question: can we build a controllable world model of the cell? Not a static embedding of cell state, but a simulator — give it a cell and an intervention, and have it predict the consequence.

The argument in three steps

  1. Predict in representation space. Biological measurements are dominated by noise: dropout above 90% in scRNA-seq, batch effects, assay artifacts. Reconstructing raw counts spends capacity on what is fundamentally unpredictable. The JEPA move — predict the latent of the held-out part, $\hat z = P(z_{\text{ctx}})$, against an EMA target — discards the noise by construction.
  1. Condition the predictor on actions. A representation is not yet a world model. The world model is the transition: $\hat z_{t+1} = P(z_t, a_t)$. In biology the action $a_t$ is an intervention — a knockout, a compound at a dose, a combination. This is the V-JEPA 2-AC recipe pointed at cells, which is exactly what BioJEPA-AC attempts.
  1. Plan in the latent space. Once transitions are learned, drug-discovery questions become search problems: which intervention moves this diseased state toward a healthy manifold while avoiding toxic regions?

The honest caveat

Representation quality and perturbation accuracy are not the same objective — the Cell-JEPA result makes this concrete, improving state reconstruction far more than effect-size estimation. And the world-model identifiability theory (LeJEPA) assumes Gaussian latents and stationary additive-noise transitions that real biology violates. So the program is empirical: latent prediction is the right prior, but it must be married to true perturbation data, causal validation, and calibrated uncertainty.

That tension — elegant prior versus messy biology — is what the rest of this site is for.