arXiv · 2611.17705
S1Pert: Scalable Target-Specific Gene Response Prediction with System-One Intelligence
Abstract
Predicting transcriptional responses to unseen genetic perturbations is essential for extending CRISPR screens across vast target--gene combinations. However, models can achieve strong aggregate predictive performance by reproducing historical gene-level response frequencies without adequately distinguishing the effects of different perturbations. Reliable prediction at the transcriptome scale is further challenged by uncertain response labels and redundant computation across millions of target--gene pairs. We present S1Pert, a System-One framework for target-specific gene response prediction with non-autoregressive online inference. S1Pert combines directional Welch tests and equivalence testing to distinguish statistically supported responses from inconclusive observations. A frozen System-One encoder extracts reusable semantic representations of perturbation targets, query genes, and historical evidence. Compact pair-specific interactions are then computed and cached offline. A lightweight hierarchical cross-attention decoder integrates these representations with cross-target pairwise ranking and synchronized historical-evidence dropout to reduce reliance on gene-level response frequencies. Extensive experiments demonstrate that S1Pert improves target-specific response discrimination while maintaining reliable prediction and efficient inference across diverse cellular contexts. These findings highlight the potential of System-One decision learning for scalable perturbation-response prediction beyond historical gene-level response tendencies. The source code is publicly available at https://github.com/liujiawen-jpg/S1Pert.
Explore related subjects
Keep this discovery
Explore connections, maps & timelines
Xiaoqi Sheng, Hou wang, Yan Liang, Yutong Li, Wenhao Du, Tinghe Zhang, Jiawen Liu, Hongmin Cai. 2026-11-13. S1Pert: Scalable Target-Specific Gene Response Prediction with System-One Intelligence. https://arxiv.science/abs/2611.17705
Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite arXiv Science by Indiciae: BibTeX · RIS.