README.md

metadata

title: MapPFN
emoji: 🧬
colorFrom: indigo
colorTo: red
sdk: gradio
sdk_version: 6.31.0
app_file: app.py
short_description: Predict single-cell gene perturbation effects in context
python_version: '3.12'
startup_duration_timeout: 30m
license: mit
models:
  - marvinsxtr/MapPFN
datasets:
  - marvinsxtr/MapPFN

MapPFN: Learning Causal Perturbation Maps in Context

Demo of MapPFN (Sextro, Kłos & Dernbach, NeurIPS 2026): a prior-data fitted network that predicts post-perturbation single-cell expression distributions in a biological context by reading other perturbation experiments from that context in-context.

• Weights: marvinsxtr/MapPFN. The demo uses model.ckpt, the checkpoint pre-trained on the synthetic prior (EMA weights). The released fine-tuned checkpoints don't match the gene panel/order of the released data splits, so they are not used.
• Data: marvinsxtr/MapPFN (split_10, 50-gene log-normalised panels).
• Code: github.com/marvinsxtr/MapPFN (MIT). The original JAX/Equinox model is ported 1:1 to PyTorch (mappfn_torch.py). Checkpoints are deserialised with Equinox on CPU at startup. Sampling matches the authors' settings: Dopri5 with a constant step of 0.01 and classifier-free guidance of 2.0.