The model sits between two kinds of public data: spatial atlases of real embryos — the measured morphology it must reproduce — and regulatory databases — the genome’s own instructions. This page describes both sides.
The atlases: MOSTA, ZESTA, HESTA
MOSTA — the mouse. The Mouse Organogenesis Spatiotemporal Transcriptomic Atlas (Stereo-seq) spans E9.5 to E16.5; every captured spot carries both its position and its transcriptome. The model reaches held-out R² = 0.70 on it, and by stacking all thirteen serial E16.5 sections we reconstructed a dense 3D fetus (208,000 cells, 30 tissues) in which the looping heart is topologically resolved for the first time — the benchmark against which the model’s forward-grown organs are scored.
ZESTA — the zebrafish. The zebrafish atlas adds the time axis cheaply — whole embryos, many stages. The model reaches R² = 0.85, and the temporal series shows the bioelectric field condensing out of the blastula (spatial coherence rising from 0 to 0.94): the model’s central object appearing in real data.
HESTA — the human. The human embryo atlas is the bridge from mouse to man: the frozen mouse compiler transfers zero-shot (the human heart scores 0.71 where an uninformed baseline scores 0.42), and a low-rank species adapter lifts the per-organ mean from 0.46 to 0.73. The human embryo is not remodelled from scratch — it is the mouse model, adapted.
The databases: AlphaGenome, SEdb, ABC
ABC — which enhancer drives which gene. The Activity-by-Contact model maps enhancers to their target genes by combining chromatin activity with 3D contact frequency. In Cognimed the ABC map is what converts the genome into physiology: summed over the ion-channel and gap-junction genes of a tissue it yields conductances, the conductances go through the Goldman equation, and out comes the tissue’s resting membrane potential — the bioelectric setpoint, computed rather than asserted.
SEdb — where the identities live. The Super-Enhancer Database catalogues hundreds of thousands of super-enhancers — the large regulatory domains that mark cell identity genes. In the model a cell type is a cluster of master-transcription-factor super-enhancers, and differentiation is attention over these clusters; SEdb is the census of the available heads.
AlphaGenome — from sequence alone. AlphaGenome is the frozen sequence front-end: given raw DNA it predicts the accessibility, expression and contact tracks that the other two databases consume. It validates and extends them — where a track is missing it can be predicted, and for an individual variant it supplies the direction of regulatory effect, which is what lets adult GWAS results be read back into the model (see the manifold view).
Together the two sides make the model a glass box — ChIP-seq checks its weights, ATAC its activations, RNA-seq its outputs, patch-clamp its conductances, voltage imaging its latent state — and the training process is the loop that reconciles them.