Researchers introduce intervention-aware distillation framework that uses gene expression data to improve drug discovery image analysis

A transcriptome-conditioned teacher model integrates gene expression and intervention metadata to guide image representation learning from microscopy, while an image-only student learns to predict these distributions independently at test time.

The framework explicitly handles dose and cell-type mismatches in weakly paired data by employing a fine-tuned single-cell foundation model to encode cell-type context and disentangle dose effects, organizing representations by drug similarity in a chemistry-aware codebook.

On Cell Painting and RxRx datasets paired with L1000 transcriptomics, the method significantly improves one-shot transfer to unseen interventions and drug-target gene discovery compared to self-supervised and alignment baselines.