Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens
A paper titled 'Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens' was published on arXiv on 2026-09-10. It introduces AssayBench-Loop, a benchmark of 1,389 CRISPR screens across five phenotype categories, and AssayLoop, a framework combining AssayFormer (a transformer-based amortized acquisition policy) with LLM-derived biological priors.
The paper addresses sequential experimental design in CRISPR screening under constrained budgets. It presents AssayBench-Loop as a large-scale benchmark for adaptive hit discovery and AssayLoop as a sequential experimental design framework. AssayLoop uses AssayFormer, a transformer-based amortized acquisition policy trained across historical screens, and incorporates LLM-derived biological priors through an adaptive handoff.
The approach treats completed experiments as training data for learning how accumulated evidence should guide next experiments. AssayFormer is a transformer-based amortized acquisition policy, suggesting a shift from per-experiment Bayesian optimization to learned policies that can generalize across screens. The integration of LLM priors via an adaptive handoff indicates a hybrid system where LLMs seed the search and learned policies refine it based on feedback.
This work signals growing interest in AI-driven experimental design for biotechnology, particularly CRISPR screening. The use of large-scale benchmarks and amortized policies could reduce the cost and time of hit discovery. The involvement of LLMs suggests a trend toward combining generative models with specialized acquisition policies in scientific workflows.
The framework could lower experimental costs and accelerate drug target discovery by prioritizing perturbations more effectively. It may create opportunities for AI-driven lab automation platforms and partnerships between AI research groups and biotech firms.
Observable next signals include: adoption of AssayBench-Loop by other research groups, publication of follow-up studies applying AssayLoop to real-world CRISPR screens, and potential commercialization of similar frameworks by biotech or AI companies. Further development may focus on expanding benchmark diversity and improving the adaptive handoff between LLMs and learned policies.