---
title: Temporal Cell State Prediction
emoji: 🔬
colorFrom: blue
colorTo: red
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
pinned: false
license: mit
short_description: Frozen V-JEPA-2 predicts cell-cycle state in microscopy
---

# Temporal Cell State Prediction in Microscopy

A frozen **V-JEPA-2 ViT-L** encoder + a small attentive-pool head predicts per-cell
cell-cycle state (interphase / pre-mitosis / mitosis) directly from short single-cell
clips — replacing the *classify* stage of the conventional segment → track → classify
pipeline with a single clip → state model.

This Space serves **pre-computed** results (no heavy model load on the free CPU tier):

1. **Classification gallery** — correct calls + the characteristic failure modes.
2. **Videos** — whole-FOV predicted-state overlay + a Trackastra-style tracking view.
3. **Counting** — the same frozen encoder probed for per-state cell counts (honest: beats a null on total, not per-state).
4. **VLM point-reasoning** — a zero-shot "point at each nucleus then sum" baseline.

**Study finding:** in this small-data regime, **data scaling, not model scaling, is the
binding constraint** — moving the baseline from a small to a larger labelled dataset lifts
macro-F1 far more than an ~80× larger encoder does.

- Models: [`DnaRnaProteins/vjepa2-cell-cycle-vit-l`](https://huggingface.co/DnaRnaProteins/vjepa2-cell-cycle-vit-l), [`DnaRnaProteins/unet-bilstm-cell-cycle-baseline`](https://huggingface.co/DnaRnaProteins/unet-bilstm-cell-cycle-baseline)
- Data: MICCAI Cell Tracking Challenge — Fluo-N2DL-HeLa. Labels derived from lineage trees (no manual annotation).