AssayLoop：结合历史实验反馈与大模型先验的自适应CRISPR筛选框架

My Thinking Process: Summarizing an Active Learning Framework for Biological Discovery

I'm focusing on how to approach biological discovery problems where experiments have to be chosen one after the other, especially when there's a limited budget. CRISPR screening is a classic example of this, because it's usually impossible to test every possible perturbation. So, we need smart ways to prioritize which perturbations to try in each round. Existing benchmarks for this adaptive hit discovery are not scaled up enough. That's why I'm introducing AssayBench-Loop. It's a large-scale benchmark for adaptive hit discovery that includes a lot of CRISPR screens across different biological traits, and makes it possible to learn about acquisition strategies from past experiments.

Building on that, I'm presenting AssayLoop, a framework that designs experiments sequentially. This framework uses AssayFormer, which is a Transformer-based system that quickly learns how to acquire new data from past experimental data. It works by combining this with biological knowledge that comes from large language models, using something I call an "adaptive handoff." Completed experiments turn into training data, helping to learn how accumulated evidence should guide the selection of the next experiment, with the language models supplying some initial biological information. There's also AssayLLM, showing how the same principle works directly with an LLM by training it specifically for the task. In tests where some of the data is set aside, AssayLoop does much better than just picking things randomly and recovers a significant portion of the hits after only a small number of tests. It's also better than other active learning methods, stand-alone LLMs, and using AssayFormer by itself. The framework performs even better with more historical training data, and can be applied to biological traits that it wasn't even trained on. These results really highlight the benefits of learning acquisition strategies from past experiments and combining them with broad biological knowledge to make adaptive hit discovery more efficient.

许多生物学发现问题都需要在预算受限的情况下序贯地选择实验。CRISPR 筛选就是一个典型的例子，因为穷举式的扰动测试往往不可行，必须在多轮实验中对候选扰动进行优先级排序。尽管该问题至关重要，但现有的自适应命中发现基准在规模和多样性上仍然十分有限。在此，我们推出了 AssayBench-Loop，这是一个用于自适应命中发现的大规模基准，包含跨越 5 个表型类别的 1389 项 CRISPR 筛选。除了支持系统性评估之外，它的规模还使得从历史实验中学习采集策略成为可能。基于这一资源，我们提出了 AssayLoop，这是一个序贯实验设计框架，它通过自适应衔接机制，将 AssayFormer（一种基于 Transformer 的摊销采集策略，通过跨历史筛选训练来根据实验反馈进行自适应调整）与源自大语言模型（LLM）的生物学先验相结合。在此视角下，已完成的实验转化为训练数据，用于学习累积的证据应如何指导下一步测试，而 LLM 则提供生物学先验知识来为搜索提供初始引导。我们进一步推出了 AssayLLM，证明了通过特定任务的后训练，同样的原理可以直接扩展到 LLM 中。在基于时间划分的留出筛选实验上，AssayLoop 相比随机选择实现了 5.67 倍的富集，并在仅测试了约 5% 的候选文库后就捕获了 27.7% 的命中项，性能优于现有的自适应设计方法、单独的 LLM 以及单独的 AssayFormer。其性能随着历史训练数据的增加而提升，并能迁移至未包含在训练中的表型类别。这些结果证明了跨历史实验学习采集策略并将其与广泛的生物学先验相结合，以实现高效自适应命中发现的价值。