iiinova — where your digital self keeps evolving

Marvin Sextro @marvinsextro.de · 30/09/2026

[3/3] Many thanks to my co-authors Weronika Kłos and Gabriel Dernbach for the great collaboration, and to @tuberlin.bsky.social, @bifold.berlin, Aignostics and Charité for their support! Looking forward to presenting our work at NeurIPS! #NeurIPS2026

000

Marvin Sextro @marvinsextro.de · 30/09/2026

[2/3] MapPFN brings prior-data fitted networks to perturbation prediction, adapting to new biological contexts and gene sets. 🌐 Website: marvinsxtr.github.io/MapPFN 📄 Paper: arxiv.org/abs/2601.21092 💻 Code: github.com/marvinsxtr/MapPFN

marvinsxtr.github.io

MapPFN: Learning Causal Perturbation Maps in Context

MapPFN is a prior-data fitted network that uses in-context learning to predict single-cell perturbation effects, pre-trained on a synthetic biological prior of in silico gene knockouts.

100

Marvin Sextro @marvinsextro.de · 30/09/2026

[1/3] Happy to share that our paper "MapPFN: Learning Causal Perturbation Maps in Context" has been accepted at NeurIPS 2026! 🎉 TL;DR: We frame perturbation prediction as an in-context learning problem given a set of interventional experiments.

100

Marvin Sextro @marvinsextro.de · 22/04/2026

[6/6] My amazing co-author Weronika Kłos will present MapPFN at the Generative AI in Genomics Workshop (Gen²) at #ICLR2026 in Rio. 🇧🇷 📅 Mon, Apr 27, 2026, 1:10–1:55 PM (UTC-3) 📍 Room 211, Riocentro Convention and Event Center

000

Marvin Sextro @marvinsextro.de · 22/04/2026

[5/6] 🌐 Project Page: marvinsxtr.github.io/MapPFN/ 📄 Paper: arxiv.org/pdf/2601.21092 💻 Code: github.com/marvinsxtr/M... Huge thanks to Weronika Kłos and Gabriel Dernbach for this collaboration, and to @tuberlin.bsky.social, @bifold.berlin , Aignostics, and Charité for their support!

marvinsxtr.github.io

MapPFN: Learning Causal Perturbation Maps in Context

MapPFN is a prior-data fitted network that uses in-context learning to predict single-cell perturbation effects, pre-trained on a synthetic biological prior of in silico gene knockouts.

111

Marvin Sextro @marvinsextro.de · 22/04/2026

[4/6] A single pre-trained MapPFN adapts to new datasets and arbitrary gene sets. Zero-shot, it recovers differentially expressed genes on par with baselines trained on real single-cell data. Fine-tuned, it consistently outperforms baselines across downstream datasets.

110

Marvin Sextro @marvinsextro.de · 22/04/2026

[3/6] MapPFN meta-learns to map pre- to post-perturbation distributions from a synthetic biological prior of in silico gene knockouts, decoupling it from limited experimental data. At inference, it adapts to unseen biological contexts via in-context learning.

110

Marvin Sextro @marvinsextro.de · 22/04/2026

[2/6] Single-cell perturbation datasets cover only a tiny slice of possible interventions and cell states. Existing methods cannot leverage new interventional evidence at inference time, forcing them to retrain for every new dataset.

220

Marvin Sextro @marvinsextro.de · 22/04/2026

[1/6] How can we build virtual cell foundation models that adapt to unseen biological contexts? Meet MapPFN, the first prior-data fitted network (PFN) for perturbation prediction. Meta-learned from a synthetic biological prior, it adapts at inference via in-context learning. 🧵

162