Can phenotypic activity be predicted without experimental readouts?

Télio Cropsal; Rocío Mercado

cs.LGOct 6, 2026

Can phenotypic activity be predicted without experimental readouts?

Authors: Télio Cropsal, Rocío Mercado

Organizations: AI Laboratory for Molecular Engineering (AIME) Department of Computer Science and Engineering Chalmers University of Technology & University of Gothenburg, Gothenburg, Sweden · AI Laboratory for Molecular Engineering (AIME) Department of Computer Science and Engineering Chalmers University of Technology & University of Gothenburg and Science for Life Laboratory (SciLifeLab), Gothenburg, Sweden

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.

Figures & tables

Figure 1: A molecular perturbation can be used two ways towards the same classification task. Experimentally (top), the compound is applied to cells and its effect is captured as a Cell Painting morphological profile; copairs [ 12 ] permutation testing on that profile separates active compounds from inactive ones, and the same test applied to cell count alone flags toxicity, giving the ground-truth label. Computationally (bottom), the same classification is predicted directly from a molecular embedding of the perturbation, or a phenotype-pretrained embedding like CLOOME or CellCLIP.

Figure 2: Train, validation, and test AUROC for CLOOME, CellCLIP, and PhysChem on BBBC036v1’s official split, target active, not toxic. Bars: mean AUROC across bootstrap resamples of the corresponding fold’s predictions; error bars: 95% CI (

n=1{,}000

resamples).

Figure 3: Representation ranking, target active, not toxic. (a) BBBC036v1 official split, restricted to the four representations with no morphology-derived pretraining (CLOOME and CellCLIP on this split are shown in Figure 2 ). (b) JUMP-CP chemical-generalization split, all six representations. (c) JUMP-CP batch-generalization split, all six representations. Bars: mean AUROC; error bars: 95% CI (bootstrap resampling in (a),

5\times 5

repeated grouped CV in (b)-(c)).

Figure 4: Representation ranking on JUMP-CP’s bioactive-library sub-data set ( source_7 ), target active, not toxic. (a) Chemical-generalization split. (b) Plate-generalization split. Bars: mean AUROC; error bars: 95% CI (

5\times 5

repeated grouped CV).

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Representation tested	Input	Dim.	Source	Model
PhysChem	SMILES	217	RDKit descriptors, deterministic	Logistic regression
LogP	SMILES	1	Crippen logP, deterministic	Logistic regression
MorganFP	SMILES	2,048	Morgan fingerprint, $r=2$	Logistic regression
CLOOME	Morgan FP, $r=3$ , 1,024 bits	512	Frozen pretrained checkpoint	Logistic regression
CellCLIP	Templated SMILES prompt	512	Frozen pretrained checkpoint	Logistic regression
MLP	Morgan FP, $r=3$ , 1,024 bits	n/a	n/a	Entire network (end-to-end)

Appendix

Table 1: Summary of every representation and model evaluated in this work for the various sets of labels. All models are

\ell_{2}

-regularized logistic regression probes. Only the MLP control is trained end-to-end; in every other row the representation is fixed and the probe is the only thing fitted.

Figure 5: Mean test AUROC by prediction target (active; active and toxic; active, not toxic) for all six representations, on (a) BBBC036v1’s official split and (b) JUMP-CP’s chemical-generalization split. Error bars: 95% CI, as in Figure 3 .

Explore similar work

Learning Molecular Representations from Cellular Phenotypes with Structure Preservation
Aug 3, 2026Xuan Lin, Jingyu Sheng, Tengfei Ma +2Molecular Representation Learning Chemical Structures
School of Computer Science, Xiangtan University, Xiangtan, Hunan, China · College of Computer Science and Electronic Engineering, Hunan University, Changsha, Hunan, China · School of Computer Science, Beijing University of Posts and Telecommunications, Beijing, China +1
Beyond Predictive Accuracy: A Reliability-Aware Audit of Molecular Representations for Human Olfaction
Jul 25, 2026Kai Lun Huang, Wei Chieh SunMolecular Representation Learning Smells
California State University, Fullerton · Department of Electrical and Computer Engineering California State University, Fullerton Fullerton, CA, USA · University of Washington +1
A systematic investigation of molecular encoding methods for drug property predictions across neural network and Transformer encoder-based model
Jun 8, 2026Sheng-Ya Chen, Shan-Ju YehMolecular Property Prediction Mechanistic Interpretability
Department of Life Science, National Tsing Hua University, Hsinchu, Taiwan. · 4Interdisciplinary Program of Life Sciences and Medicine, National Tsing Hua University, Hsinchu, Taiwan. · School of Medicine, National Tsing Hua University, Hsinchu, Taiwan. +1

Aug 3, 2026, cs.LGJ/K move · Enter open · S save

Learning Molecular Representations from Cellular Phenotypes with Structure Preservation

Xuan Lin, Jingyu Sheng, Tengfei Ma, Li Sun+1

School of Computer Science, Xiangtan University, Xiangtan, Hunan, China · College of Computer Science and Electronic Engineering, Hunan University, Changsha, Hunan, China · School of Computer Science, Beijing University of Posts and Telecommunications, Beijing, China+1

Molecular Representation Learning Chemical Structures Cells Cross-Modal Alignment Preservation

Open paper AskPDF arXiv

Phenotypic drug discovery enables the discovery of functional relationships between molecular structures and cellular responses. However, existing multimodal representation learning methods often optimize cross-modal alignment without considering the intrinsic organization of chemical space, resulting in distorted molecular representations and loss of structural information. We propose \textbf{PhenMol}, a structure-preserving framework for phenotype-aware molecular representation learning. PhenMol disentangles molecular and cellular representations into shared and private components, enabling phenotype-guided alignment while preserving chemical structures through a dedicated molecular branch. This design integrates cellular phenotype information without disrupting molecular neighborhood organization. Experiments on approximately

3.04 \times 10^{4}

molecule--cell morphology pairs demonstrate that PhenMol improves molecular property prediction across 270 bioactivity tasks, molecule--phenotype retrieval, and clinical trial outcome prediction. Moreover, ECFP4-based structural analysis shows that PhenMol better preserves molecular neighborhoods and reduces embedding distortion compared with existing multimodal alignment methods. These results highlight the importance of structure-aware constraints in multimodal molecular representation learning and provide an effective approach for integrating cellular phenotypes with chemical knowledge for drug discovery.

Abstract

Explore similar work

Learning Molecular Representations from Cellular Phenotypes with Structure Preservation

Beyond Predictive Accuracy: A Reliability-Aware Audit of Molecular Representations for Human Olfaction

A systematic investigation of molecular encoding methods for drug property predictions across neural network and Transformer encoder-based model