Molecular encoders do not outperform simple descriptors in predicting cell traits
Can phenotypic activity be predicted without experimental readouts?
Machine Learning
Summary
Scientists wanted to find out if computer models trained on images of cells could predict how molecules affect cell behavior without doing lab experiments. They checked several models carefully, controlling for tricky issues that can make results seem better than they are. They found that these fancy models didn't predict cell traits any better than simple chemical features of the molecules. Predicting if a molecule is toxic was easier than predicting more complex changes in cells. The authors suggest being careful when trusting these models to replace lab tests in drug discovery.
What this means in practice
- •For drug discovery teams: Avoid relying solely on pretrained molecular encoders to predict phenotypic effects and continue using experimental assays for more accurate results.
- •For chemical safety assessors: Use simple chemical descriptors as reliable indicators when screening for molecule toxicity since they perform as well as complex pretrained models.
Authors
Télio Cropsal, Rocío Mercado
Abstract
Molecular encoders contrastively pretrained on paired molecule-morphology data, such as CLOOME and CellCLIP, have been proposed as cheap surrogates for phenotypic prediction, avoiding the need to run a Cell Painting assay. We evaluate this idea for these molecular encoders under a protocol designed to control for two confounds that can inflate apparent performance: leakage across an encoder's own pretraining boundary, and the correlation between phenotypic activity and cytotoxicity. Testing six representations, including a non-pretrained MLP control matching CLOOME's input and layer count, on two distinct Cell Painting screens, we find that once these confounds are controlled for, the pretrained molecular encoders show no clear advantage over plain physicochemical descriptors, and that toxicity is generally easier to predict than phenotypic activity across representations. Our results suggest leakage-aware, confound-controlled evaluation should be standard practice before phenotype-pretrained encoders are trusted as surrogates for phenotypic drug discovery.