Do Perturbation Predictors Learn Gene Interactions? Diagnosing Extrapolation in Agentic Screening
Abstract
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. 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. 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. 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. We further study how this signal can guide decisions in an agentic perturbation-screening workflow. The code is available at https://github.com/zygao930/OOD.
Related
- High-throughput platforms for genetic perturbation screening using CRISPR/Cas9 in human iPSC-derived macrophages for drug discovery
- STEER-seq: a recombinase based platform for pooled multimodal single-cell perturbation screening
- Genotype × Isolate Interactions for Tan Spot Resistance Screening in Mungbean
- Automated screening of autism via a pretrained vision-language model in naturalistic caregiver-child interactions.
- Performances and Critical Limitations of Docking and Protein–Ligand Interactions for Efficacy-Driven Virtual Screening Targeting Class A GPCRs
BibTeX
@misc{gao2026perturbation,
title = {{Do Perturbation Predictors Learn Gene Interactions? Diagnosing Extrapolation in Agentic Screening}},
author = {Ziyuan Gao},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.18880},
eprint = {2610.18880},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.18880},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Do Perturbation Predictors Learn Gene Interactions? Diagnosing Extrapolation in Agentic Screening
AU - Ziyuan Gao
PY - 2026
DA - 2026/10/01
PB - arXiv
DO - 10.48550/arxiv.2610.18880
UR - https://arxiv.org/abs/2610.18880
AB - 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. 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. 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. 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. We further study how this signal can guide decisions in an agentic perturbation-screening workflow. The code is available at https://github.com/zygao930/OOD.
ER -