Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens

Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens introduces AssayBench-Loop, a large-scale benchmark comprising 1,389 CRISPR screens across five phenotype categories, designed for sequential, budget-constrained experiment selection. Building on this resource, the paper presents AssayLoop, a framework that combines AssayFormer (a transformer-based amortized acquisition policy) with LLM-derived biological priors to iteratively prioritize candidate perturbations.

On temporally held-out screens, AssayLoop achieves a 5.67-fold enrichment over random selection and recovers 27.7% of hits after assaying approximately 5% of the candidate library, outperforming existing adaptive-design methods and standalone LLMs. The framework demonstrates that learning acquisition policies across historical experiments and integrating broad biological priors significantly enhances efficiency in adaptive hit discovery.

Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens introduces AssayBench-Loop, a benchmark for studying experiment selection in adaptive biological discovery pipelines. The central setting is sequential, budget-constrained hit discovery, such as CRISPR screening, where a researcher or algorithm must decide which assays to run next based on partial evidence, limited budget, and the goal of identifying biologically meaningful hits. Rather than treating screening as a one-shot classification or ranking task, the work frames it as an interactive decision-making problem in which model predictions, experimental outcomes, and future assay choices are tightly coupled.

The key contribution is a large-scale benchmark for evaluating amortized adaptive methods in this domain. “Amortized” here points to the need to make sequential experimental decisions efficiently, reusing learned structure or representations across rounds instead of incurring prohibitive retraining or inference costs at each step. By providing a common benchmark, the material enables systematic comparison of algorithms that balance exploration and exploitation, update beliefs from noisy biological measurements, and allocate limited screening budget to maximize hit discovery. This is particularly relevant for CRISPR screens, where assay costs, guide complexity, and biological uncertainty can make naive exhaustive screening impractical.

The work matters because it connects active learning, sequential experimental design, and biological discovery in a setting where real experiments are expensive and iterative. A standardized benchmark can help move the field beyond isolated case studies by making it easier to assess whether adaptive policies genuinely improve hit discovery efficiency, robustness, and reproducibility. More broadly, AssayBench-Loop supports the emerging vision of “biology-in-the-loop” systems, where machine-learning-guided experiment selection is integrated with laboratory workflows to accelerate discovery while reducing wasted screening effort.