EMBER: Contextualized gene representations enable compositional prediction across cellular perturbations

EMBER, a gene-centric foundation model, learns context-dependent gene representations that improve cross-context perturbation-response prediction and genetic-interaction ranking versus evaluated baselines.

• Why it matters: Predicting effects of single and combined perturbations is needed to find context-specific vulnerabilities, but transferring gene representations across perturbation tasks and cellular states is difficult.
• What they did: EMBER pretrains on a diverse single-cell corpus using interaction-guided contrastive learning plus masked expression reconstruction, then freezes the encoder and trains lightweight probes for composable prediction.
• The result: The pretrained embeddings support single-gene, digenic interaction, and multi-target drug-response prediction, enabling better cross-context performance and prioritizing SCFD1 as a candidate dependency during mTOR inhibition.

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