Local-Atlas Control-Anchored Flow Matching for Unpaired Single-Cell Perturbation Prediction

Destructive single-cell perturbation assays ob- serve control and treated populations without true before/after pairs, so response prediction is an unpaired conditional distribution problem rather than paired regression. We introduce Atlas-CFM, a local-atlas control-anchored flow- matching formulation for this setting. The method constructs training paths and inference rollouts from the same observed control-derived anchor, uses transport-aware pseudo-pairing to reduce tra- jectory noise, and regularizes dynamics with a soft tangent/normal vector-field decomposition. On Norman perturbation-OOD and sciPlex3-K562 dose-extrapolation benchmarks, the largest gains occur in the harder generalization regimes. Sup- plementary native-space checks, distribution met- rics, consistency studies, sensitivity analyses, and A549 cross-context results show that training- inference-consistent control anchoring strength- ens perturbation-response modules in life-science foundation-model systems.