PerturbMap: Cross-Context Transfer of Single-Cell Perturbation Responses

TL;DR - PerturbMap predicts missing single-cell perturbation responses by transferring evidence across cellular contexts while weighting routes by validation-estimated reliability. On a melanoma dataset, it improved prediction accuracy and preserved context specificity better than several baselines.

• Combines a recipient-specific low-rank model with transported responses from source contexts.
• Uses ridge experts trained on paired perturbations and weights proposals according to validation-anchor reliability.
• Reduced full-effect MSE by 4.1% versus the recipient-local low-rank baseline.
• Increased top-10 same-recipient counterpart retrieval by cosine similarity from 74.5% to 80.5%.