Pre-Existing Heterogeneity Predicts Rare Proteostasis-Stress Programs Across Diverse Perturbations

Stress-associated transcriptional programs are common in single-cell perturbation data, but they are often treated as technical or experimental nuisance signals. Whether rare high-stress populations arise stochastically after perturbation or reflect outcomes associated with pre-existing cellular heterogeneity remains unclear. Herein, we build a cross-dataset stress-program prediction framework spanning 146,321 single cells and 926 perturbation–cell-line tasks. Untreated baseline heterogeneity, together with perturbation identity, predicted future rare integrated-stress burden across held-out cell-line–drug pairs (R2=0.742, Pearson r=0.862). Single-cell stress-program prediction generalized across leave-task-out, leave-cell-line-out and leave-perturbation-out splits; retained signal in leave-dataset-out evaluation; and collapsed to near-null performance under within-task label permutation. Independent validation datasets provided external support for the inferred stress axes: tunicamycin and thapsigargin activated unfolded protein response/integrated stress response (UPR/ISR) modules in bulk RNA sequencing (RNA-seq), thapsigargin expanded populations with high X-box binding protein 1 (XBP1) UPR or activating transcription factor 4 (ATF4) ISR activity in donor-paired single-cell data, and an independent protocol-stress dataset indicated heat shock/proteostasis structure beyond cell type and quality control (QC). The resulting resource summarizes perturbation responses as activator protein 1 (AP1) immediate, UPR/ATF4, heat shock, replication-coupled and low/mixed dominant stress-program patterns. These findings suggest that stress variation should not be viewed solely as a nuisance covariate but may represent a predictable and biologically structured dimension of perturbation response space.