Learning single-cell perturbation responses using neural optimal transport.

Area of Science:

• Single-cell biology
• Computational biology
• Systems biology

Background:

• Predicting cellular responses to perturbations is crucial but challenging due to destructive single-cell measurement techniques.
• Existing methods struggle with unpaired data from perturbed and non-perturbed cells, limiting understanding of heterogeneous responses.

Purpose of the Study:

• To develop a novel computational framework, CellOT, for predicting individual cell responses to perturbations.
• To address the challenge of unpaired data distributions in single-cell analysis.

Main Methods:

• Leveraging optimal transport theory and input convex neural networks.
• Developing a framework (CellOT) to map unpaired distributions of perturbed and non-perturbed single cells.
• Validating predictions using scRNA-seq and multiplexed protein-imaging data.

Main Results:

• CellOT significantly outperforms existing methods in predicting single-cell drug responses.
• Demonstrated generalization by predicting responses in holdout patient data (lupus, glioblastoma) and across species (lipopolysaccharide response).
• Successfully modeled hematopoietic developmental trajectories.

Conclusions:

• CellOT provides a powerful new approach for analyzing single-cell perturbation responses.
• The framework enhances predictive accuracy and generalizability across diverse biological contexts.
• CellOT facilitates deeper understanding of cellular heterogeneity and responses to various stimuli.