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

UnDetect predicts drug-induced transcriptional responses to transcriptionally unprofiled compounds on a drug-blind L1000 benchmark, reaching 0.117 versus 0.092 for the strongest published model.

• Why it matters: Drug-blind prediction is limited because most models underuse the compound signal, so removing drug input barely changes accuracy and compound-specific effects are missed.
• What they did: UnDetect separates contextual variation from compound effects using a normalized multi-view kernel that integrates chemical structure, annotated targets, mechanisms of action, and Cell Painting morphology, then transfers residual responses within and across profiled cell lines.
• The result: On drug-blind L1000 it beats a compound-ignoring network in every fold, and on DeepCE it achieves 0.602 Pearson correlation versus 0.491; transfer improves 28/28 LINCS L1000 and 47/50 and 49/50 Tahoe-100M evaluations at higher doses.

AI summary of the abstract · may contain errors · not medical advice · read the paper.