Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens
Exhaustive testing in CRISPR screens remains economically prohibitive, demanding efficient sequential search policies. To address this, researchers introduced AssayBench-Loop, a benchmark spanning 1,389 screens across five phenotype classes, alongside AssayLoop, an adaptive experimental design framework. AssayLoop initiates target selection using LLM-derived biological priors and dynamically hands off decisions to AssayFormer—a transformer policy trained across historical screens to learn from feedback. On temporally held-out screens, testing approximately 5% of the candidate library recovers 27.7% of hits, achieving a 5.67-fold enrichment over random search while generalizing to unseen phenotype categories.
Why it's worth reading
It provides a clear template for closing the lab-in-the-loop discovery cycle, combining broad LLM prior knowledge with amortized learning across historical wet-lab screens.