Separating context from drug effects improves expression prediction for unseen compounds
- Existing models of drug-induced expression changes turned out to rely very little on the compound itself; UnDetect separates shared context from compound-specific effects.
- It fuses structure, targets, mechanism and Cell Painting information in a kernel tolerant of missing data, and transfers residual responses across cell lines.
- On drug-blind L1000 and DeepCE benchmarks it outperformed published models, and cross-cell-line transfer helped in nearly all LINCS and Tahoe-100M evaluations.
💡 Points Based on the abstract
If earlier models barely used the compound, a context-only predictor is the baseline that matters, so I'd look at how much gain survives over it when test compounds are scaffold-distinct, not just drug-blind. L1000 signatures are noisy, so average correlations can flatter; I'd want replicate-level ceilings. On my project, would it rank close analogs of one series correctly?