Research map / Probabilistic and scientific modeling / Causal discovery

MapPFN: Learning Causal Perturbation Maps in Context

NeurIPS 2026 · Probabilistic and scientific modeling › Causal discovery · approach: Latent-variable causal estimation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces MapPFN, a prior-data fitted network pre-trained on synthetic SCM and gene-regulatory biological priors that predicts single-cell perturbation distributions in context from observational and interventional data, zero-shot or after fine-tuning.

Problem

Predicting single-cell perturbation responses in unseen biological contexts and for arbitrary gene sets is limited by scarce perturbation datasets, and existing models cannot use new interventional evidence at inference time.

Approach

Meta-learn a perturbation-distribution map with a PFN: a MMDiT-style diffusion/flow transformer conditioned on observational cells and a set of interventional experiments in context, pre-trained purely on synthetic causal priors (linear SCMs and SERGIO-simulated GRNs with counterfactual paired noise), so it adapts to new contexts via in-context learning without retraining.

Key results

• Zero-shot MapPFN (pre-trained) on melanoma reaches AUPRC 0.34 and MR 1.00, comparable to CPA/CondOT/MFM range and below STATE's 0.33 only marginally; fine-tuned MapPFN reaches AUPRC 0.38 (best) with MR 0.99.
• On the leukemia dataset, fine-tuned MapPFN achieves AUPRC 0.18 and best PDS (0.42), MR (0.91) and RMSE among methods, while pre-trained-only MapPFN overshoots (MR 2.56).
• In the linear SCM benchmark, MapPFN achieves MR 0.99 versus 0.09 for CondOT and 0.12 for MFM, with the best MMD, RMSE and PDS.

Surprising findings

• Pre-trained-only MapPFN overshoots the ground-truth effect on leukemia (MR 2.56) and has very high W2 there, which the authors attribute to the dataset containing only 26 perturbation targets.
• A randomly initialized MapPFN trained only on real data is competitive on W2 and leukemia AUPRC, performing worse than fine-tuned MapPFN on most other metrics.

Limitations

• Synthetic prior models only hard CRISPR knockouts, so soft CRISPRi knockdowns and drug perturbations are not covered.
• Pre-trained-only performance is inconsistent across datasets; fine-tuning is needed on the leukemia data.
• Model is small (about 25M parameters) and the authors note scaling model capacity is future work.

Experimental setup

Datasets: Melanoma Perturb-CITE-seq (Papalexi/Liu-style CRISPR screen, ~218k cells) THP-1 leukemia ECCITE-seq (~20k cells) synthetic linear SCM prior synthetic SERGIO biological prior

Benchmarks: Linear SCM prior benchmark Melanoma CRISPR perturbation dataset Leukemia THP-1 ECCITE-seq dataset

Metrics: W2 (Sinkhorn Wasserstein) MMD RMSE PDS magnitude ratio AUPRC

Baselines: CPA CondOT Meta Flow Matching CellFlow STATE Identity Observed

Builds on: prior-data fitted networks Multimodal Diffusion Transformer (MMDiT) conditional flow matching SERGIO CellFlow

Closest papers by meaning

• Identifying biological perturbation targets through causal differential networks (ICML 2025 · similarity 0.90)
• Generative Intervention Models for Causal Perturbation Modeling (ICML 2025 · similarity 0.89)
• Do-PFN: In-Context Learning for Causal Effect Estimation (NeurIPS 2025 · similarity 0.88)
• Contextualizing biological perturbation experiments through language (ICLR 2025 · similarity 0.87)
• PerturBench: Benchmarking Machine Learning Models for Cellular Perturbation Analysis (NeurIPS 2025 · similarity 0.87)
• Learning Identifiable Factorized Causal Representations of Cellular Responses (NeurIPS 2024 · similarity 0.86)