MapPFN: Causal Perturbation Prediction

• MapPFN is a perturbation prediction model that leverages meta-learning and synthetic causal priors to predict post-intervention single-cell distributions.
• It employs a prior-data fitted network and in-context learning, conditioning on both untreated and interventional data to adapt in new biological settings.
• Empirical results indicate that MapPFN effectively recovers differentially expressed genes, demonstrating robust zero-shot performance and improvements upon fine-tuning.

Searching arXiv for the MapPFN paper and closely related context. MapPFN is a perturbation prediction model for biological intervention settings, especially single-cell CRISPR perturbation data, in which the objective is to predict how a population of cells changes under an intervention in a new biological context. It is introduced as a prior-data fitted network (PFN) that meta-learns a causal perturbation map from synthetic data generated under a prior over causal perturbation mechanisms, then performs amortized inference in context at test time by conditioning on observational and interventional evidence from the target context rather than retraining for that context (Sextro et al., 28 Jan 2026). In the formulation given for single-cell perturbation modeling, the task is distribution-to-distribution prediction because cells are unpaired: one observes untreated single-cell expression samples and some measured intervention outcomes in the same target context, and seeks the post-perturbation distribution for an unseen intervention (Sextro et al., 28 Jan 2026).

1. Problem setting and causal target

MapPFN is motivated by the fact that perturbation effects are context-dependent: the same gene knockout can produce different outcomes in different cell lines, states, or microenvironmental conditions because the underlying causal mechanisms differ. In single-cell assays, experiments are expensive, destructive, and sparse across contexts, and cells are measured only once, so paired pre/post states for the same cell are unavailable. The modeling problem is therefore not sample-wise prediction but population-level prediction of a post-perturbation distribution from an observational distribution together with any interventional evidence already collected in the same context (Sextro et al., 28 Jan 2026).

The paper formalizes the observed untreated population as

$Y^{\mathrm{obs}} \in \mathbb{R}^{N \times d},$

where $N$ is the number of cells and $d$ the number of genes. For an intervention $t \in \mathcal{T}$, the target post-intervention population is

$Y^{\mathrm{int}} \in \mathbb{R}^{M \times d}.$

Conditioning is performed not only on the untreated population but also on a context set of measured intervention experiments,

$C = \{(t_k, Y_k^{\mathrm{int}})\}_{k=1}^{K},$

with $t_k \in \mathcal{T}_C \subset \mathcal{T}$. The prediction target for an unseen query intervention $t_q \in \mathcal{T}\setminus \mathcal{T}_C$ is

$$p(y_q^{\mathrm{int}} \mid do(t_q), Y^{\mathrm{obs}}, C). \tag{1}$$

This is the causal perturbation map that MapPFN is trained to approximate (Sextro et al., 28 Jan 2026).

The causal interpretation assumes that each biological context is generated by a latent structural causal model (SCM) representing a gene regulatory network (GRN), with single-node hard interventions corresponding to gene knockouts and all variables of the latent SCM observed. Under these assumptions, the posterior predictive distribution is written as

$$p(y_q^{\mathrm{int}} \mid do(t_q), Y^{\mathrm{obs}}, C) = \int p(y_q^{\mathrm{int}} \mid do(t_q), Y^{\mathrm{obs}}, \mathcal{M}) \, p(\mathcal{M} \mid Y^{\mathrm{obs}}, C)\, d\mathcal{M}. \tag{2}$$

This framing makes explicit that prediction marginalizes over uncertainty in the underlying causal mechanism, conditioned on both observational and interventional evidence. MapPFN does not explicitly recover the causal graph; instead, it amortizes this inference by pretraining across many synthetic mechanisms $N$0 (Sextro et al., 28 Jan 2026).

2. Prior-data fitted network formulation

Conceptually, MapPFN is a PFN for perturbation modeling. PFNs are pretrained on synthetic datasets sampled from a prior over data-generating mechanisms and then perform amortized Bayesian inference via in-context learning at inference time. In MapPFN, the hypothesis space is a prior over causal perturbation mechanisms rather than a prior over labels, scalar targets, or per-sample outputs. The paper emphasizes that the target is a distribution of vectors, which distinguishes the setting from standard PFNs for tabular prediction or causal inference (Sextro et al., 28 Jan 2026).

The operational representation of biological context at inference time is the combination of the untreated observational distribution $N$1 and the small interventional context set $N$2. A perturbation is represented by a treatment identifier $N$3, usually a one-hot encoded gene knockout, augmented in synthetic SCMs by intervention strength and represented in real data as gene identity. The output is a predicted post-perturbation cell-level distribution,

$N$4

that is, a generated set of synthetic cells representing the outcome distribution for the query intervention (Sextro et al., 28 Jan 2026).

Inference-time adaptation is achieved through in-context learning. The model is not retrained from scratch for a new biological context. Instead, newly observed interventions from that context are inserted into the context set $N$5, and the pretrained network conditions on them to infer the likely perturbation map. The paper therefore characterizes the method as transductive and amortized: adaptation is performed by conditioning rather than optimization. The associated causal intuition is that interventional data shrink the interventional Markov equivalence class of the latent causal graph, so additional context interventions help determine which mechanisms are compatible with the target context (Sextro et al., 28 Jan 2026).

Empirically, the paper reports that increasing context size $N$6 improves performance monotonically, and increasing the number of cells per contextual intervention at test time also improves performance. This is presented as evidence of genuine in-context adaptation rather than the behavior of a fixed predictor that ignores adaptation inputs (Sextro et al., 28 Jan 2026).

3. Synthetic causal priors and pretraining procedure

A central contribution of MapPFN is its synthetic pretraining regime. During each pretraining step, the model samples a mechanism $N$7 from a prior over causal systems. In the controlled SCM benchmark, $N$8 is a linear additive noise model over variables $N$9 with DAG $d$0 and structural assignments

$d$1

represented in the linear case by a sparse weighted adjacency matrix

$d$2

With Gaussian noise $d$3, samples are generated via

$d$4

A hard intervention $d$5 on node $d$6 removes incoming edges and sets $d$7, yielding $d$8 (Sextro et al., 28 Jan 2026).

For biological pretraining, the prior is made more realistic by replacing simple linear SCMs with synthetic GRNs. These GRNs are scale-free, modular, and sparse, sampled via preferential attachment and simulated with SERGIO. SERGIO uses Hill-function regulatory interactions and stochastic differential equations to generate realistic single-cell gene expression, and in silico gene knockouts are implemented by removing the perturbed gene from the regulatory network and re-simulating. Technical noise matching 10x Chromium scRNA-seq is then added. This synthetic biological prior is intended to cover realistic perturbation mechanisms while using only simulated data (Sextro et al., 28 Jan 2026).

The pretraining loop consists of sampling a mechanism $d$9, generating observational samples $t \in \mathcal{T}$0, sampling a context set of interventions $t \in \mathcal{T}$1, generating context post-perturbation populations $t \in \mathcal{T}$2, selecting an unseen query intervention $t \in \mathcal{T}$3, and generating the corresponding target $t \in \mathcal{T}$4. The network is trained to predict $t \in \mathcal{T}$5 from $t \in \mathcal{T}$6. The synthetic data therefore encode a prior over causal perturbation maps directly (Sextro et al., 28 Jan 2026).

An important pretraining design choice is the “counterfactual paired prior”: during synthetic pretraining with SERGIO, the same random seed is kept across treatments so that differences between intervention distributions are attributable to the intervention mechanism rather than stochastic initial-condition variation. The paper states that this isolates causal signal and materially improves downstream performance. Table 4 and Figure 1 are reported to show that removing this paired prior hurts all metrics and leads to slower and worse convergence (Sextro et al., 28 Jan 2026).

4. Architecture and generative objective

MapPFN uses a Multimodal Diffusion Transformer (MMDiT) adapted from Stable Diffusion 3. The architectural motivation is that the output is a distribution of vectors rather than a scalar estimate. Cells are treated as tokens, and the model contains three modality streams with separate parameters: noisy target samples, cell state inputs, and treatment embeddings. Cross-modal interaction occurs through joint attention (Sextro et al., 28 Jan 2026).

Because the cell sets are unordered, sinusoidal positional encodings are removed and the model relies on attention’s permutation invariance over set inputs. To preserve semantic distinctions among inputs, the network adds learnable embeddings that indicate different treatment conditions in context, observational versus interventional data, and query versus context. The architecture has 8 layers, embedding dimension 256, feed-forward expansion to 512, 4 attention heads of size 64, and 8 register tokens appended to the noise stream, totaling about 25M parameters. Time conditioning is implemented using FiLM (Sextro et al., 28 Jan 2026).

The generative objective is conditional flow matching. The network learns a velocity field $t \in \mathcal{T}$7 that transports Gaussian noise toward the target post-perturbation distribution. The loss is

$t \in \mathcal{T}$8

where

$t \in \mathcal{T}$9

with $Y^{\mathrm{int}} \in \mathbb{R}^{M \times d}.$0 and $Y^{\mathrm{int}} \in \mathbb{R}^{M \times d}.$1. During inference, the learned flow is integrated as an ODE using Dopri5, with classifier-free guidance and default weight $Y^{\mathrm{int}} \in \mathbb{R}^{M \times d}.$2 for conditional generation (Sextro et al., 28 Jan 2026).

The full training pipeline has three stages. The first is synthetic pretraining: 50k steps on linear SCMs or 400k steps on the biological prior. Training uses AdamW, a warmup-stable-decay schedule, exponential moving average with decay 0.999, and condition dropout $Y^{\mathrm{int}} \in \mathbb{R}^{M \times d}.$3 in which conditions are replaced by a learned null embedding. In the biological prior, the paper simulates 6000 synthetic GRNs of 50 genes with $Y^{\mathrm{int}} \in \mathbb{R}^{M \times d}.$4 cells per treatment and context size $Y^{\mathrm{int}} \in \mathbb{R}^{M \times d}.$5. The second stage is optional fine-tuning on real single-cell perturbation data for 3,000 iterations with linear warmup to learning rate $Y^{\mathrm{int}} \in \mathbb{R}^{M \times d}.$6, reported to take about 10 minutes on one GPU. The third stage is ODE-based sampling at inference time (Sextro et al., 28 Jan 2026).

5. Benchmarks, datasets, and evaluation

The baselines listed in the paper are CPA, CondOT, Meta Flow Matching (MFM), CellFlow, and STATE, together with identity and oracle references. These methods are positioned by conditioning capability: CPA, CondOT, and CellFlow condition on covariates; MFM and STATE additionally use observational populations; only MapPFN is described as conditioning on interventional populations at test time. Table 1 is reported to summarize this distinction (Sextro et al., 28 Jan 2026).

The evaluation covers three datasets or benchmarks. The first is a synthetic benchmark of linear SCMs with $Y^{\mathrm{int}} \in \mathbb{R}^{M \times d}.$7 variables and context size $Y^{\mathrm{int}} \in \mathbb{R}^{M \times d}.$8. The second is a melanoma CRISPR Perturb-CITE-seq dataset from Frangieh et al. with about 218k cells, 248 gene perturbations, and three biological contexts, using the IFN-$Y^{\mathrm{int}} \in \mathbb{R}^{M \times d}.$9-treated cell line as holdout context. The third is a leukemia ECCITE-seq CRISPR dataset from Papalexi et al. with about 20k cells and 26 perturbations in one context. For real-data experiments, the study restricts to 50 genes, using perturbed genes plus top markers, and samples $C = \{(t_k, Y_k^{\mathrm{int}})\}_{k=1}^{K},$0 cells per condition. Evaluation follows the Virtual Cell Challenge setup: for the held-out context, some interventions are available in training as adaptation context and the rest are held out for testing (Sextro et al., 28 Jan 2026).

The reported metrics span several aspects of quality: entropy-regularized Wasserstein distance / Sinkhorn divergence, MMD, RMSE of mean expression, perturbation discrimination score (PDS), magnitude ratio (MR), and AUPRC for differentially expressed gene recovery. The magnitude ratio is introduced to normalize for intervention effect scale,

$C = \{(t_k, Y_k^{\mathrm{int}})\}_{k=1}^{K},$1

with $C = \{(t_k, Y_k^{\mathrm{int}})\}_{k=1}^{K},$2 a distributional distance such as Wasserstein. According to the paper, $C = \{(t_k, Y_k^{\mathrm{int}})\}_{k=1}^{K},$3 indicates that the predicted intervention effect has the correct magnitude relative to the untreated state, whereas $C = \{(t_k, Y_k^{\mathrm{int}})\}_{k=1}^{K},$4 indicates collapse to identity. This metric is emphasized because some models achieve small raw distances by barely moving away from the untreated distribution (Sextro et al., 28 Jan 2026).

6. Empirical results and ablations

On synthetic linear SCMs, MapPFN is reported as best across all reported metrics. The paper gives the following values: $C = \{(t_k, Y_k^{\mathrm{int}})\}_{k=1}^{K},$5, $C = \{(t_k, Y_k^{\mathrm{int}})\}_{k=1}^{K},$6, $C = \{(t_k, Y_k^{\mathrm{int}})\}_{k=1}^{K},$7, $C = \{(t_k, Y_k^{\mathrm{int}})\}_{k=1}^{K},$8, and $C = \{(t_k, Y_k^{\mathrm{int}})\}_{k=1}^{K},$9. By contrast, CondOT and MFM are described as having MR near 0.1, indicating near-identity collapse. The paper interprets Table 2 as strong evidence that conditioning on interventional context and using the PFN setup helps recover the true intervention effect (Sextro et al., 28 Jan 2026).

On real single-cell data, the most prominent result is that a single MapPFN pretrained only on synthetic in silico knockouts can recover differentially expressed genes in zero-shot mode at a level comparable to models trained directly on real data. On the melanoma dataset, zero-shot pretrained MapPFN reaches $t_k \in \mathcal{T}_C \subset \mathcal{T}$0, comparable to CondOT at $t_k \in \mathcal{T}_C \subset \mathcal{T}$1 and STATE at $t_k \in \mathcal{T}_C \subset \mathcal{T}$2, while achieving the best $t_k \in \mathcal{T}_C \subset \mathcal{T}$3 at $t_k \in \mathcal{T}_C \subset \mathcal{T}$4. On the leukemia dataset, zero-shot distributional performance is worse, which the paper attributes to weaker alignment with the synthetic prior, but AUPRC remains competitive at $t_k \in \mathcal{T}_C \subset \mathcal{T}$5, roughly similar to baselines in the range $t_k \in \mathcal{T}_C \subset \mathcal{T}$6–$t_k \in \mathcal{T}_C \subset \mathcal{T}$7 (Sextro et al., 28 Jan 2026).

After fine-tuning, MapPFN is reported to improve consistently and generally outperform all baselines. On melanoma, it reaches $t_k \in \mathcal{T}_C \subset \mathcal{T}$8, best $t_k \in \mathcal{T}_C \subset \mathcal{T}$9, and $t_q \in \mathcal{T}\setminus \mathcal{T}_C$0. On leukemia, it reaches best $t_q \in \mathcal{T}\setminus \mathcal{T}_C$1, best $t_q \in \mathcal{T}\setminus \mathcal{T}_C$2, best $t_q \in \mathcal{T}\setminus \mathcal{T}_C$3, best $t_q \in \mathcal{T}\setminus \mathcal{T}_C$4, and best $t_q \in \mathcal{T}\setminus \mathcal{T}_C$5, though not best $t_q \in \mathcal{T}\setminus \mathcal{T}_C$6. The paper also states that fine-tuning a pretrained model yields better performance on nearly all metrics than training MapPFN from random initialization, which is presented as evidence that synthetic meta-learning remains useful after exposure to real data (Sextro et al., 28 Jan 2026).

Two ablation findings are highlighted. First, removing interventional context hurts all metrics, which supports the central claim that the model benefits specifically from adaptation using observed interventions in the target context. Second, removing the counterfactual paired prior also hurts all metrics. Figure 1 is further described as showing faster and better convergence with the paired prior. Appendix Figure 2 is said to show improvements from more context experiments and more cells at inference time, reinforcing the in-context learning interpretation. Figure 3 reports test-time augmentation to 100 genes using overlapping 50-gene subsets, which improves AUPRC and reduces variance (Sextro et al., 28 Jan 2026).

7. Scope, strengths, and limitations

The paper identifies several strengths. It presents MapPFN as the first PFN for perturbation prediction; a method that performs context-adaptive inference using observed interventions from the target context; a framework that decouples pretraining from the scarcity of real perturbation datasets by using synthetic biological priors; an approach that supports arbitrary gene sets via test-time augmentation; and a model that shows strong zero-shot and fine-tuned performance, especially on DEG recovery and effect magnitude. It is also framed as having a principled causal formulation and as exploiting interventional information in a manner that most baselines do not (Sextro et al., 28 Jan 2026).

The stated limitations are equally explicit. Synthetic-to-real transfer depends on the quality and coverage of the synthetic biological prior, and mismatch harms zero-shot performance, as illustrated by the leukemia dataset with bimodal expression distribution. The method assumes that transfer across contexts is possible because the relevant real mechanisms lie within the support of the synthetic prior. It supports only single-gene hard interventions akin to CRISPR knockouts, not soft knockdowns, combinatorial perturbations, drug treatments, or chemical perturbations. It assumes that all variables of the latent SCM are observed, which the paper notes is unrealistic in many biological settings. Adaptation to larger gene sets currently relies on augmentation over subsets rather than direct high-dimensional modeling. Pretraining incurs a one-time cost of roughly 10–36 GPU hours. Finally, some reported benefits depend on the counterfactual paired prior, whose broader validity and desirability are left open (Sextro et al., 28 Jan 2026).

In practical terms, MapPFN is intended for settings in which one wants to predict post-perturbation single-cell distributions in a new biological context and has access to an untreated population from that context, a small set of observed intervention populations from the same context, and a query intervention identity. The output is a generated population of post-perturbation cells from which one can compute distributional summaries, perturbation-specific signatures, and differentially expressed genes. This distinguishes it both from standard perturbation prediction models, which usually learn a global predictor over training contexts and condition only on treatment and perhaps baseline expression, and from standard treatment-effect estimation, because the target is a full high-dimensional population distribution under intervention rather than a scalar average treatment effect (Sextro et al., 28 Jan 2026).

A plausible implication is that MapPFN occupies a distinct methodological niche: it treats biological perturbation prediction as amortized causal inference over distributions with explicit use of interventional context at test time. Within the evidence reported in the paper, its defining empirical claim is that synthetic-pretrained zero-shot models already recover DEGs competitively with real-data-trained baselines, while fine-tuning yields state-of-the-art downstream performance on the tested single-cell datasets (Sextro et al., 28 Jan 2026).