Can phenotypic activity be predicted without experimental readouts?
The pretrained molecule encoders lost their edge once leakage and cytotoxicity were controlled.
The paper tests CLOOME, CellCLIP, and other representations on two Cell Painting screens under a stricter evaluation protocol. A plain physicochemical-descriptor baseline and even a non-pretrained MLP control were competitive once the authors guarded against pretraining-boundary leakage. Toxicity was generally easier to predict than phenotypic activity, which the authors flag as a confound in prior-looking results. They argue these encoders should not be treated as assay substitutes without leakage-aware, confound-controlled testing. ArXiv · AI/CL/LG's note
The paper tests CLOOME, CellCLIP, and other representations on two Cell Painting screens under a stricter evaluation protocol. A plain physicochemical-descriptor baseline and even a non-pretrained MLP control were competitive once the authors guarded against pretraining-boundary leakage. Toxicity was generally easier to predict than phenotypic activity, which the authors flag as a confound in prior-looking results. They argue these encoders should not be treated as assay substitutes without leakage-aware, confound-controlled testing. ArXiv · AI/CL/LG's note
score 4