Intervention-Aware AI Framework for Drug Discovery from Microscopy

A novel artificial intelligence framework has been introduced in a preprint on arXiv (2604.22832), enhancing drug discovery by merging microscopy-based phenotypic profiling with transcriptomic data. This technique, known as intervention-aware distillation, employs a teacher-student model. The teacher, conditioned on transcriptomic data, synthesizes gene expression and intervention metadata to generate soft distributions based on a chemistry-aware codebook aligned by drug similarity. Utilizing a fine-tuned single-cell foundation model, the teacher captures cell-type context and differentiates dose effects. Meanwhile, an image-only student predicts these distributions solely from microscopy images, extracting mechanistic insights without the need for expensive transcriptomic assays. This method overcomes the shortcomings of current multimodal approaches that overlook cell-type and dose variations in weakly paired datasets. The announcement was made on April 28, 2026.