Can phenotypic activity be predicted without experimental readouts?

2026-10-06

AI Summary

This study investigates whether molecular encoders can effectively predict phenotypic activity in the absence of experimental readouts, noting that existing evaluations are frequently compromised by data leakage and confounding cytotoxicity. To address this, the authors propose a rigorous evaluation protocol that systematically controls for these confounders, benchmarking pretrained encoders against conventional molecular descriptors on Cell Painting data. The findings reveal that, upon eliminating such confounders, current encoders offer no significant advantage over basic descriptors. Furthermore, while general cytotoxicity proves readily predictable, specific phenotypic responses remain exceedingly challenging to forecast. Consequently, this work advocates establishing leakage-proof evaluation as a standard practice within the field.

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.

Problem

• phenotypic prediction
• molecular encoders
• data leakage
• cytotoxicity confound
• Cell Painting

Innovation

• Molecular Encoders
• Phenotypic Prediction
• Data Leakage
• Confound-controlled Evaluation
• Cell Painting