Do Perturbation Predictors Learn Gene Interactions? Diagnosing Extrapolation in Agentic Screening

High-throughput perturbation screens increasingly use machine-learning models to predict the effects of gene combinations that were never measured. A difficult case arises when both genes in a test pair were seen individually during training, but their combination was not.

高通量扰动筛选越来越多地使用机器学习模型来预测从未测量过的基因组合效应。当测试对中的两个基因在训练过程中都曾单独出现，但它们的组合未曾出现时，就会出现一个难题。

In this setting, standard OOD methods often fail because neither gene appears unfamiliar on its own. We therefore compare each model prediction with a simple additive prediction obtained by summing the two observed single-gene effects.

在这种情况下，标准的分布外（OOD）方法往往会失效，因为单独来看，没有任何一个基因显得陌生。因此，我们将每个模型的预测结果与通过将两个已观测到的单基因效应相加而获得的简单加性预测进行了比较。

The difference between them, which we call additive disagreement (AD), measures how much extra interaction effect the model introduces for an unseen combination. Across eight prediction methods and three Perturb-seq datasets, specialised neural predictors do not consistently outperform this simple additive baseline.

两者之间的差异，我们称之为加性分歧（AD），用于衡量模型为未见过的组合引入了多少额外的相互作用效应。在八种预测方法和三个 Perturb-seq 数据集上，专门的神经预测器并没有始终优于这一简单的加性基准。

More importantly, predictions with larger deviations from the additive baseline tend to have larger errors. Conventional output-based and embedding-based confidence signals are inconsistent across datasets and models, whereas AD remains informative across all evaluated compositional settings.

更重要的是，与加性基准偏差越大的预测，往往误差也越大。传统的基于输出和基于嵌入的置信度信号在不同数据集和模型间表现不一致，而 AD 在所有评估的组合设置中始终保持信息量。

We further study how this signal can guide decisions in an agentic perturbation-screening workflow. The code is available at this https URL.

我们进一步研究了该信号如何指导代理扰动筛选工作流程中的决策。代码可在该网址获取。