A modular causal adaptation architecture for machine learning on data-scarce rare disease.

Rare diseases are called orphan diseases because each affects few patients, leaving too little data to model. Perturbation foundation models condition on experimental context via a learned lookup table, so a context never seen during pretraining has no embedding and no prediction, and a patient-derived line carrying a particular lesion will never be screened at that scale. Rather than learning rare-disease biology from scratch, I isolate the pathological shift within a pre-wired model of general biology, parameterizing the mutation as a rank-one deviation added to a frozen structural causal model. This reduces the estimation problem from quadratic to linear in the number of genes, permitting inference from roughly ten observations, with a soft one-hot vector identifying the network entry node. A modular loop adds counterfactual scoring, bootstrap-ensemble uncertainty, and active experiment selection. On simulated networks of 50 nodes, localization accuracy rises from 0.00 to 0.72 as observations increase from 5 to 40, against a chance level of 0.02, while an unconstrained sparse deviation reaches only 0.11. On real protein-signaling data, recovery remains at chance; controlled experiments injecting known lesions indicate that network size, not the method, limits recovery.