Biologists often cannot test every possible genetic perturbation in the lab. CRISPR screens ask which genes or edits change a cell’s behavior, but testing every candidate is too slow and expensive. This paper introduces tools that use past experiments and language models to choose which perturbations to try next, so researchers can find important “hits” with far fewer tests.

The authors created AssayBench-Loop, a large benchmark made from 1,389 CRISPR screens that cover five different kinds of observable effects (phenotypes). They also built AssayLoop, a system for choosing experiments in rounds. AssayLoop pairs AssayFormer, a transformer-based model that learns a general strategy from many past screens, with biological background supplied by large language models (LLMs). They also present AssayLLM, which applies the same idea by fine-tuning an LLM directly for this task.

At a high level the method treats finished experiments as training examples. AssayFormer is an “amortized acquisition policy,” meaning it learns once from historical data how to use partial results to pick the next tests. The LLMs provide prior biological knowledge to start the search in a sensible place. An adaptive handoff lets the system combine the LLM’s suggestions with evidence collected during the new screen and update what to test next.

On screens that were held out in time from the training data, AssayLoop gave concrete gains. It achieved a 5.67-fold enrichment of hits compared with picking candidates at random. After testing roughly 5% of the candidate library, the method recovered about 27.7% of the true hits. The paper reports that AssayLoop outperformed prior adaptive-design methods, standalone LLMs, and AssayFormer acting alone. The authors also found that the method got better when it had more historical screens to learn from, and that it could transfer to phenotype categories that were not in its training set.