Predicting Transcriptional Responses to Unseen Compounds through Context Modeling and Multi-View Transfer

Predicting drug-induced transcriptional responses could support in silico screening of new compounds, yet existing models barely use the compound: on a drug-blind L1000 benchmark, removing the drug input from seven published models changes accuracy by at most 0.012 Pearson correlation. They capture shared cellular and experimental variation while overlooking the compound-specific signals that distinguish drug effects. We present UnDetect, a framework that separates contextual variation from compound effects to predict responses to transcriptionally unprofiled compounds. UnDetect integrates chemical structure, annotated targets, mechanisms of action and Cell Painting morphology into a normalized multi-view kernel that accommodates incomplete compound information, and transfers residual responses within and across profiled cell lines, so that pharmacological and phenotypic relationships inform predictions beyond chemical similarity. On a drug-blind benchmark, where the paired control reveals the assay plate, UnDetect is the only model to beat a network that ignores the compound, in every fold; on a within-plate score that is zero for any compound-blind prediction, it reaches 0.117 against 0.092 for the strongest published model. On DeepCE's benchmark it reaches 0.602 Pearson correlation against the reported 0.491. Transfer across cell lines improves 28 of 28 LINCS L1000 and, at the two higher doses, 47 and 49 of 50 Tahoe-100M cell-line evaluations. These results show that separating context from compound effects and integrating complementary compound information improves transcriptional prediction and extends the utility of perturbation atlases for virtual screening.