Reproduce “DRIFT: Disentangled Responsive-Invariant Flow Transport for Single-Cell Perturbation Prediction” on OpenResearch

About this paper

Mustapha Bounoua, Giulio Franzese, Pietro Michiardi

DRIFT predicts single-cell gene-expression responses to genetic and chemical perturbations by separating each cell’s latent state into invariant and responsive components, then transporting only the responsive component with a conditional flow model. It outperforms reported baselines on most metrics across three benchmarks, including held-out perturbation combinations and unseen perturbations, while results do not establish individual-cell causal counterfactuals.

Key insights

• Conditioning response transport on an invariant cell representation allows predictions to vary with baseline cell state without deliberately transporting that state.
• Ablations identify disentanglement as the largest overall contributor to performance; invariance penalties generally help, and perturbation-feature regularization is especially useful for unseen conditions.
• The latent diagnostics are consistent with the intended separation, but they do not prove that the learned components are uniquely identifiable biological factors.

Summary by alphaXiv.