Genetic perturbation data lets a model forecast drug transcriptional responses in cell types never treated with the drug
- GRAFT transfers knowledge from genetic perturbation responses to chemical ones and across cellular contexts, predicting drug-induced expression changes without drug data in the target cell type.
- On LINCS and external benchmarks it outperformed competing methods, notably for context-dependent drugs.
- An atlas covering many contexts and thousands of drugs was generated, with docking support for some associations and partial support for donor-specific iPSC differences from measured genetic perturbations.
💡 Points Based on the abstract
Knockdown-to-drug transfer quietly assumes a compound mimics loss of its target, which should fray for off-target or agonist-like effects. Docking agreement is also weak validation of an expression forecast. A fair test of the atlas is to hold out one cell type and run your own compounds, ideally ones with known off-target activity.
Cross-modal transfer enables zero-shot prediction of drug responses across cell types ↗