PerturbPFN: Probing the Limits of Synthetic Priors in Drug Perturbation Modelling

Predicting how cells respond to novel small-molecule drugs is a fundamental challenge in drug discovery. Existing models often rely on black-box statistical learning or dataset-specific optimization, which struggle to generalize to unseen perturbations or provide interpretable insights into the underlying biological mechanisms. The authors ask whether a model can be trained entirely on synthetic data to perform 'in-context' learning, effectively inferring latent causal structures and intervention effects for new, unseen drug treatments.

The authors introduce PerturbPFN, a framework inspired by Prior-Data Fitted Networks (PFNs). Instead of directly regressing high-dimensional gene expression data, the model uses a hierarchical synthetic prior to learn a latent structural causal model (SCM). During training, the model is exposed to synthetic episodes generated by biologically motivated simulators that define gene regulatory networks, sparse intervention targets, and intervention strengths. At inference time, the model takes context observations (known drug-response pairs) and a query drug, then performs a single forward pass to predict the latent system graph, the specific genes targeted by the drug, and the magnitude of the effect. These estimates are then propagated through an SCM decoder to predict the final cellular response.

PerturbPFN demonstrates competitive performance on real-world single-cell perturbation datasets compared to specialized baselines, while maintaining low inference costs. Because it is trained on synthetic episodes, it avoids the need for expensive, dataset-specific fine-tuning. Beyond predicting expression outcomes, the model provides interpretable intermediate estimates, such as the inferred gene regulatory graph and the specific atomic targets of a drug, which are typically unavailable in standard end-to-end deep learning approaches.