Spectral Adaptive Repositioning for Flow-Based Single-Cell Perturbation Modeling

SPEAR predicts transcriptome-wide responses to genetic, pharmacological, and cytokine perturbations by adaptively repositioning genes using perturbation-conditioned hidden states across layers.

• Why it matters: Current models rely on arbitrary gene orderings or static co-expression masks from unperturbed cells, missing functional rewiring caused by perturbations.
• What they did: SPEAR combines spectral positional encoding from graph Laplacian geometry with adaptive repositioning that updates each gene’s position from its perturbation-conditioned hidden state at every layer.
• The result: Across perturbation benchmarks, SPEAR shows excellent recovery of differential gene-expression, distributional structure, and perturbation-specific identifiability, and it learns latent positional relationships matching known transcription factor–target interactions.

AI summary of the abstract · may contain errors · not medical advice · read the paper.