ScreenShot: A Foundation Model for Few-Shot Combination Drug Screening

TL;DR - ScreenShot is a hierarchical transformer that predicts combination-drug responses from a few observations on a new patient, without molecular profiling or fine-tuning. It could reduce screening costs while improving personalized treatment selection.

• Pretrained on 40 datasets spanning 3,700 drugs and 6,000 biological samples.
• Uses in-context learning over functional measurements to make few-shot predictions.
• Outperformed baselines on four held-out datasets in accuracy and selective-treatment identification.
• Its active-learning strategy matched uniform screening’s hit detection using one-third of the budget.