High-content microscopy enables systematic profiling of cellular responses to chemical perturbations, but the scale of the chemical space makes exhaustive phenotypic characterization experimentally infeasible. This motivates computational models that can predict image-derived phenotypes without acquiring the corresponding treated cells. We formulate molecule-induced phenotype prediction as an inductive conditional transport problem in image representation space. Given a negative-control phenotype and the structure of a molecule, we aim to predict the phenotype induced by the corresponding molecule. We first evaluate classical optimal transport baselines and show that static couplings do not yield useful predictions on large-scale phenotypic image datasets. We then introduce a molecule-conditioned Neural Optimal Transport (NOT) model with a Monge-Gap regularization training objective that learns to transport negative-control unperturbed phenotypes toward perturbed phenotypes using molecular structure as conditioning information. NOT recovers molecule-specific phenotypic effects while reducing microscopy-associated technical variation, thereby facilitating comparisons across experimental batches. On unseen active molecules, the model outperforms baseline approaches, demonstrating that chemically conditioned transport can generalize beyond the molecules observed during training. We identified the molecular encoder as the main limitation to this generalization, while transport in a compressed representation space improves performance and scalability. These results establish NOT as a promising framework for predicting cellular phenotypes from molecular structure and negative-control phenotypes, while highlighting the development of more informative molecular representations as a key direction for improving out-of-distribution performance.
Figures & tables
Figure 1: Overview of the molecule-conditioned Neural Optimal Transport framework. (a) Model architecture . DMSO images and molecular structures are encoded separately and projected to a common dimension d , then combined through cross-attention to predict perturbation-induced phenotype embeddings. (b) Training objective . The conditional map Tθ(⋅∣f) transforms the source distribution μ into Tθ#μ . Learning minimizes a fitting loss between the predicted and target distributions, together with a Monge-Gap regularizer that constrains the transport. (c) Evaluation . Predictions are ranked against candidate target phenotypes by cosine similarity. Retrieval is successful when the correct perturbation appears among the top k candidates, using held-out plates for in-distribution evaluation and held-out molecules for out-of-distribution evaluation.
Figure 2: In-distribution prediction on held-out plates. (a) Retrieval@1 for the eight shared positive controls, reported as mean ± standard deviation across five folds and three seeds. (b) UMAP of source, target, and predicted phenotypes for the four most active controls. The projection was fitted jointly on source and target embeddings.
Figure 3: Generated phenotypes improve cross-plate comparability. Cross-plate retrieval performance on held-out plates. Bars show the mean mAP across controls, while points denote individual molecules. Real phenotypes are evaluated either directly or after per-plate DMSO-based normalization. Generated phenotypes achieve higher cross-plate mAP than both real baselines.
Figure 4: Neural OT on the molecule-level OOD split (5-fold, 3-seed mean ± std). (a) Replicate- and plate-level Retrieval@10 for NOT (ours) and the identity mapping: the conditioned map clearly outperforms the baseline. (b) The same metrics for four molecular structure encoders under an otherwise identical NOT, ordered by plate-level retrieval: retrieval improves substantially with encoder quality. Bars show the mean and error bars the standard deviation over five folds and three seeds. The dashed line marks the chance level ( k/M=10/2248 ).
Table 1: Impact of dimensionality reduction on OOD retrieval (5-fold, 3-seed mean ± std; best in bold, second underlined). (a) Reducing the conditioning embedding sharply hurts retrieval. (b) Reducing the phenotype embedding helps, the 50D autoencoder being best; this ablation was run at τ=0.95 , the near-optimal unbalancedness of Table 10 (Appendix 0.C.4 ).
Table 2: Ablations on the Monge-Gap strength and the conditioning mechanism (5-fold, 3-seed mean ± std; best in bold, second underlined). (a) Monge-Gap strength λ : the geometry-preserving penalty trades replicate for plate retrieval. (b) Injecting the molecular structure by multi-head attention versus plain concatenation.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 5: Examples of images from the eight positive controls of JUMP-CP.
Figure 6: Distribution of molecule-level phenotypic distances to DMSO negative controls. The red vertical line indicates the threshold separating the top ( 10% ) most phenotypically active molecules, retained for the out-of-distribution evaluation, from the remaining molecules.
Table 7: Hyperparameters for the molecule-level OOD ablation (Section 5.2 ). The settings that differ from the in-distribution setup of Appendix 0.C.2 are the data split, the entropic scales εfit,εreg , the Monge-gap weight λ , the number of epochs, and the structure encoder.
Structure encoder
R@10 rep
R@10 plate
morganc+rdkc
0.035 ± 0.002
0.095 ± 0.007
morgan
0.028 ± 0.001
0.071 ± 0.007
molformer
0.017 ± 0.001
0.040 ± 0.003
unimol2
0.015 ± 0.001
0.029 ± 0.003
Appendix
Table 8: Structure encoders under NOT on the molecule-level OOD split (5-fold, 3-seed mean ± std). Best in bold, second best underlined. These are the values plotted in Figure 4(b) .
Objective / cost
R@10 rep
R@10 plate
NOT (cosine cost)
0.035 ± 0.002
0.095 ± 0.007
NOT ( ℓ2 cost)
0.038 ± 0.002
0.095 ± 0.006
Appendix
Table 9: Ablation on the training objective and ground cost (5-fold, 3-seed mean ± std). Best in bold, second best underlined.
τ (unbalanced resampler)
R@10 rep
R@10 plate
1.00
0.035 ± 0.002
0.095 ± 0.007
0.95
0.037 ± 0.002
0.101 ± 0.007
0.90
0.037 ± 0.001
0.101 ± 0.007
Appendix
Table 10: Unbalancedness τ of the resampling heuristic (5-fold, 3-seed mean ± std). Best in bold, second best underlined. Mild unbalancedness ( τ≈0.90 – 0.95 ) gives small, consistent plate gains over the balanced τ=1 baseline.
Method
Phenotype distance
# correct (out of 1125)
GW
ℓ2
0
GW
cosine
1
Random coupling
—
1
Semi-supervised FGW ( 80% labels, λ=100 )
cosine
3
Appendix
Table 11: Static GW/FGW on the most-active JUMP-CP subset ( 1125 molecules): number of correct structure → phenotype matches. Even 80% of supervised correspondences leaves the coupling at chance level.
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 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 3.04×104 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.
Xuan Lin, Jingyu Sheng, Tengfei Ma +2
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 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.
Télio Cropsal, Rocío Mercado
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
Microscopy-based phenotypic profiling is scalable for drug discovery but lacks the mechanistic depth of transcriptomics, which remains costly and scarce. Existing multimodal approaches either use images to support other modalities or naively align representations by sample identity, ignoring cell-type and dose variations in weakly paired data-limiting generalization to unseen interventions. In this paper, we introduce an intervention-aware distillation framework that leverages perturbational transcriptomics to guide image representation learning. A transcriptome-conditioned teacher integrates gene expression and intervention metadata to produce soft distributions over a chemistry-aware codebook organized by drug similarity. The teacher employs a fine-tuned single-cell foundation model to encode cell-type context and disentangle dose effects. An image-only student learns to predict these distributions from microscopy alone, distilling mechanistic knowledge while operating independently at test time. This design emphasizes intervention semantics rather than identity alignment and explicitly handles dose and cell-type mismatches. We provide theoretical guarantees showing that transcriptomic guidance tightens the risk bound for image-based prediction. On Cell Painting and RxRx datasets paired with L1000, our method significantly improves one-shot transfer to unseen interventions and drug-target gene discovery compared to self-supervised and alignment baselines.