Single-cell perturbation profiling reveals widespread transcriptional drug escape and its dependence on baseline cell state

When a population of cancer cells is exposed to a drug, single-cell profiling reveals that a large fraction of cells fail to mount any measurable transcriptional response, remaining indistinguishable from untreated controls. We refer to these cells as \textit{transcriptional escapers} and study them systematically across two single-cell chemical-transcriptomic datasets and a clinical cohort. In sci-Plex (354,640 cells; three cancer lines; 188 compounds at four doses), escape is pervasive, on average 84.1\% (A549), 85.3\% (K562), and 84.6\% (MCF7) of treated cells, structured by drug mechanism, and dose-dependent. Using optimal-transport reconstruction of each treated cell's pre-treatment state and a leakage-controlled, held-out-drug evaluation, escape is predictable above chance in every line (mean held-out AUROC 0.827 \ensuremath{\pm} 0.003 across five random seeds), far beyond a cell-cycle-only model (0.537 \ensuremath{\pm} 0.004); the baseline programs predisposing to escape are coherent within a line but largely cell-line-specific (cross-line signature correlation \ensuremath{\leq} 0.19). On the independent MIX-Seq platform, escape replicates, quantitatively recovers drug selectivity, tracks TP53 status for the MDM2 inhibitor nutlin (Spearman \ensuremath{\rho} = \ensuremath{-}0.536 \ensuremath{\pm} 0.002), and is associated with reduced drug sensitivity for mechanism-engaging agents but not for compounds that kill without a transcriptional signature. In an exploratory analysis of a melanoma checkpoint-blockade cohort, the baseline escaper fraction among tumor-infiltrating CD8 T cells was higher in non-responders than responders (0.43 vs. 0.26; p = 0.047), suggesting the phenomenon may extend to clinical drug response, although this analysis is presently underpowered. Together these results establish transcriptional escape as a widespread, mechanism-structured, partly baseline-encoded phenomenon whose determinants are context-specific.