scCBGM: Interpretable Single-Cell Counterfactual Editing with Concept Bottlenecks and Flow Matching
The scCBGM framework leverages concept bottleneck architectures with decoder skip connections and a cross-covariance penalty for interpretable single-cell counterfactual editing. Extended to flow matching models, it allows concept-guided manipulation in both encoding-decoding and generation tasks. A synthetic benchmark aids evaluation, and scCBGM shows strong performance in generalization and prediction on real-world data.