YC
We built the first world model of the cell — and, we believe, the first to unify all four of its axes into one shared state.
Core trick
Forces expression, morphology, spatial neighbors, and time into one 128-dimensional latent state, validated against shuffled-control baselines for each axis.
Tags
Description
We built the first world model of the cell — and, we believe, the first to unify all four of its axes into one shared state. Expression, morphology, spatial context, and time have always been measured in separate experiments, on separate platforms, in separate distributions; no one had made them a single thing. We forced all four into one latent state you can read, roll forward in time, intervene on, and decode back into an image. This is what today's virtual-cell field lacks: STATE, TranscriptFormer, scGPT, GeneFlow, and Spatia are essentially single-modality transcriptomics predicting perturbation → expression — none unifies four axes, and none is a true world model. And we proved each axis is real, not an artifact, with a negative (shuffle) control on every one: expression reconstruction R² = 0.57; real spatial neighbors beat shuffled at z = −181 (p < 0.001); expression recovers morphology at r ≈ 0.66–0.68. We even report our failure honestly — the sharpest generated images were hallucinations, so we kept the coarse-but-correct one. Once four axes share one state, it becomes a universal adapter: reconstruct a modality a dataset never measured, transfer across platforms, run in-silico perturbations, and scale from 200k cells to 100M-cell atlases.
README excerpt
<h1 align="center">Cell-State World Model</h1> <p align="center"> <a href="LICENSE"><img src="https://img.shields.io/badge/license-MIT-08C?style=flat" alt="MIT" /></a> <img src="https://img.shields.io/badge/PyTorch%202.1%2B-A100%20on%20Modal-4493F8?style=flat" alt="PyTorch 2.1+, A100 on Modal" /> <img src="https://img.shields.io/badge/10x%20Xenium-200k%20cells%20%C2%B7%20422%20genes-4493F8?style=flat" alt="10x Xenium, 200k cells, 422 genes" /> <img src="https://img.shields.io/badge/workbench-React%2018%20%C2%B7%20Vite%20%C2%B7%20Claude-4493F8?style=flat" alt="Workbench: React 18, Vite, Claude" /> </p> <p align="center"> <strong>One state to draw them all.</strong><br/> A single trained network in which one 128-dimensional latent state <strong>S</strong> is shared across expression,<br/> morphology, spatial context and time. You can read S out to any axis, roll it forward, intervene on it, and<br/> decode it back into a cell image. Every axis is validated against a shuffle / identity control. </p> <h3 align="center"><a href="#getting-started"><ins>Getting started</ins></a> · <a href="report/unified_paper_full.pdf">Read the technical report (PDF)</a></h3> <p align="center"> <img src="results/figures/readme_worldmodel_overview.png" alt="Overview: expression, spatial GNN, morphology and an EMT time transfer feed one 128-d state S, which predicts all four axes; real versus generated cell morphology; the EMT trajectory in S; and the four-axis metrics against s