cs.LGSep 24, 2026

Improving Molecular-Morphology Contrastive Pretraining using Deep-Learning-based Morphology Profiles

Authors: Jie Li, Kathryn E. Kirchoff, Dante A. Pertusi, Zhizhuo Zhang

Organizations: GSK, Artificial Intelligence and Machine Learning · GSK, Cheminformatics

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Recent advancements in image-based profiling techniques have enabled the collection of high-volume cell morphology data, allowing new molecular embedding models to learn from the experimental phenotypic perturbations of a molecule in a cell. Previously, we developed Molecule-Morphology Contrastive Pretraining (MoCoP), a strategy for aligning small molecule embeddings to morphology fingerprints extracted through CellProfiler. The resulting molecular representation showed transferable performance for quantitative structure--activity relationship (QSAR) prediction tasks. Here, we extend the method by using a deep-learning-based cell image encoding pipeline to extract more feature-rich morphology profiles and align them to the molecular embeddings through contrastive learning. The new embeddings encode more accurate information on how molecules perturb cell morphology and enable improvements for QSAR predictions through either fixed-embedding linear probes or fully flexible fine-tuning. Morphology retrieval performance scales log-linearly with training data size, suggesting continued improvements as larger datasets become available. The improved MoCoP v2 also achieves superior performance on toxicity prediction and competitive results on ADME and activity benchmarks, when compared with existing molecular embedding models that use both cell morphology and transcriptomic data during training.

Figures & tables

Figure 1: Overview of the MoCoP v2 pipeline. (a) DL embedding extraction: raw 5-channel Cell Painting images undergo shading correction and tile extraction ( 224×224224\times 224 px), followed by training and featurization with a ResNet-18 model fine-tuned using triplet loss. The resulting tile-level embeddings are mean-aggregated per well to produce 128-dimensional DL morphological embeddings, which are then batch-corrected using CORAL alignment based on DMSO negative controls. (b) Contrastive pretraining: batch-corrected DL embeddings are encoded by a morphology multi-layer perceptron network (MLP), while molecular graphs are encoded by a gated graph neural network (GGNN). Both branches are projected into a shared 128-dimensional space, and the encoders are jointly trained with a symmetric InfoNCE contrastive loss. (c) Downstream applications: the pretrained GGNN molecular encoder is transferred to new molecules for QSAR prediction, ADMET prediction, molecular property prediction, mechanism-of-action retrieval, and other tasks.
Figure 2: Scaling behavior of MoCoP v2 morphology retrieval. Models are trained from scratch with 10K, 20K, 50K, and 98K compounds and evaluated on the same held-out test set. (a) 1:100 retrieval (100-candidate pool). (b) 1:1000 retrieval (1,000-candidate pool). Retrieval accuracy improves consistently with training data size across all top- kk metrics, following an approximately log-linear trend.
Figure 3: Representative examples of strong divergence between MoCoP v1 and v2 embeddings. Cell Painting composite: Blue = DNA, Green = AGP, Red = Mitochondria. Chemical structures are shown alongside each compound. Plate barcodes and well positions from JUMP-CP are also marked for reference. (a) v1 overestimates similarity for a pair with distinct phenotypes; v2 correctly identifies them as dissimilar. (b) v2 recognizes shared phenotypic effects between structurally dissimilar compounds that v1 misses.
Figure 4: Quantitative comparison of MoCoP v2, v1, and ECFP4 embedding quality, measured by CellProfiler profile retrieval on 5,265 JUMP test compounds with ≥4\geq 4 replicate wells. (a) Retrieval quality expressed as percentage of the oracle ceiling (see text), averaged over 50 random replicate splits. Error bars indicate bootstrap standard deviation. ECFP4 dominates at small kk but declines as structural analogs are exhausted; v2 overtakes ECFP4 at k≥10k\geq 10 and consistently outperforms v1 at all kk values. (b) v2 provides the best morphology retrieval for chemically isolated compounds (Tanimoto <0.3<0.3 ), exactly the regime where structure-based methods have the least information.
DataAUROCAUPRC
ProportionMoCoP v1MoCoP v2MoCoP v1MoCoP v2
5%0.570±0.0250.570\pm 0.0250.605±0.022\mathbf{0.605\pm 0.022}0.552±0.0360.552\pm 0.0360.576±0.028\mathbf{0.576\pm 0.028}
25%0.637±0.0090.637\pm 0.0090.668±0.032\mathbf{0.668\pm 0.032}0.602±0.0270.602\pm 0.0270.631±0.037\mathbf{0.631\pm 0.037}
100%0.706±0.0410.706\pm 0.0410.731±0.026\mathbf{0.731\pm 0.026}0.669±0.0440.669\pm 0.0440.695±0.039\mathbf{0.695\pm 0.039}
Table 1: Linear probe QSAR prediction performance on ChEMBL20. Mean ±\pm standard deviation across 3 splits. Bold indicates the better result.
DataAUROCAUPRC
ProportionMoCoP v1MoCoP v2MoCoP v1MoCoP v2
5%0.621±0.022\mathbf{0.621\pm 0.022}0.598±0.0150.598\pm 0.0150.569±0.0230.569\pm 0.0230.575±0.025\mathbf{0.575\pm 0.025}
25%0.689±0.018\mathbf{0.689\pm 0.018}0.677±0.0130.677\pm 0.0130.640±0.0310.640\pm 0.0310.648±0.024\mathbf{0.648\pm 0.024}
100%0.721±0.0200.721\pm 0.0200.734±0.034\mathbf{0.734\pm 0.034}0.681±0.0330.681\pm 0.0330.713±0.040\mathbf{0.713\pm 0.040}
Table 2: End-to-end fine-tuning QSAR prediction performance on ChEMBL20. Mean ±\pm standard deviation across 3 splits. Bold indicates the better result.
MethodChEMBL 2K AUC ↑\uparrowToxCast AUC ↑\uparrowBroad 6K AUC ↑\uparrowBiogen 3K MAE ↓\downarrow
Structure only
MLP76.8±2.276.8\pm 2.257.6±1.057.6\pm 1.063.3±0.363.3\pm 0.366.2±2.466.2\pm 2.4
MolT5 [ 9 ]69.9±0.869.9\pm 0.864.7±0.964.7\pm 0.955.1±0.955.1\pm 0.965.1±0.565.1\pm 0.5
UniMol [ 62 ]76.8±0.476.8\pm 0.464.6±0.264.6\pm 0.265.4±0.165.4\pm 0.155.8±2.855.8\pm 2.8
Structure + cell morphology
CLOOME [ 47 ]66.7±1.866.7\pm 1.854.2±0.954.2\pm 0.961.7±0.461.7\pm 0.464.3±0.464.3\pm 0.4
Table 3: Performance comparison on downstream benchmark tasks. Classification tasks (ChEMBL 2K, ToxCast, Broad 6K) report average AUC (%, ↑\uparrow ). Biogen 3K reports average MAE ( ×100\times 100 , ↓\downarrow ). Bold indicates the best result; underline indicates the second best. Methods are grouped by training data modalities. Values for baselines, CLOOME, InfoAlign, and CHMR are taken from [ 31 , 28 ] .
TaskMoCoP v2CHMR [ 28 ]InfoAlign [ 31 ]
LOG HLM CLint35.4±1.035.4\pm 1.033.7±0.4\mathbf{33.7\pm 0.4}39.7±0.439.7\pm 0.4
LOG MDR1-MDCK ER33.8±0.3\mathbf{33.8\pm 0.3}35.2±0.235.2\pm 0.239.2±0.339.2\pm 0.3
LOG Solubility @ pH 6.837.4±0.537.4\pm 0.534.9±0.5\mathbf{34.9\pm 0.5}40.5±0.640.5\pm 0.6
LOG PPB Human60.6±2.160.6\pm 2.153.1±1.3\mathbf{53.1\pm 1.3}66.7±1.766.7\pm 1.7
LOG PPB Rat55.1±1.155.1\pm 1.148.5±0.9\mathbf{48.5\pm 0.9}62.0±1.562.0\pm 1.5
LOG RLM CLint38.8±0.4\mathbf{38.8\pm 0.4}39.8±0.339.8\pm 0.348.4±0.648.4\pm 0.6
Table 4: Biogen 3K per-task comparison (MAE ×100\times 100 , ↓\downarrow ). Bold indicates the best result.
Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

HyperparameterValue
BackboneResNet-18 (ImageNet pretrained)
Projection headMLP: backbone→1024→128\text{backbone}\to 1024\to 128
Embedding dimension128
Tiles per well12
AggregationMean
Loss functionTriplet margin loss (margin = 0.1)
Appendix
Table 5: DL morphological embedding image encoder training hyperparameters.
StageDatasetWellsCompoundsPlates
DL image encoder training
 JUMP-CP Source 366,47155,336217
Contrastive pretraining (DL embeddings)
 Source 3 only66,47155,336217
 Full JUMP-CP (10 sources)649,384109,1641,685
Compound splits (split 0)
Appendix
Table 6: Summary of datasets used in the MoCoP v2 pipeline. “Wells” refers to the number of compound-treated wells (excluding DMSO controls used only for CORAL correction). “Compounds” counts unique chemical entities (by canonical SMILES).
Figure 5: Effect of CORAL batch correction on DL morphological embeddings. UMAP projections (cosine distance, nneighbors=30n_{\text{neighbors}}=30 ) colored by JUMP-CP data source, subsampled to ∼\sim 1,500 wells per source. (a) Before CORAL correction: embeddings cluster by data source. Batch effects dominate the clustering. (b) After CORAL correction using DMSO negative controls: source-specific clusters dissolve, and embeddings are intermixed across sources. Samples with outlier embedding norms ( << 1% of data) were excluded from the corrected panel.
HyperparameterValue
Molecular encoder (GGNN)
Input atom features75
Edge types1
Message-passing layers6
FC dimensions1024
Dropout0.1
Appendix
Table 7: MoCoP v2 contrastive pretraining hyperparameters.
Figure 6: Additional strong divergence: v2-similar pairs. MoCoP v2 correctly identifies phenotypic similarity not captured by v1.
Figure 7: Additional strong divergence: v1-similar pairs. MoCoP v1 incorrectly assigns high similarity (cos >0.5>0.5 ) to compound pairs with very different phenotypes. v2 correctly identifies these as dissimilar.
Figure 8: Divergence examples among structurally dissimilar compounds (Tanimoto <0.2<0.2 ), imaged on the same microscope (Source 3). (a–b) V2-similar: v2 correctly identifies phenotypic similarity that v1 misses, demonstrating the value of DL-embedding-aligned representations for identifying shared biology across distinct scaffolds. (c–d) V1-similar: v1 incorrectly assigns high similarity to pairs with visibly different phenotypes; v2’s lower score is more appropriate.
DatasetTask Type# Compounds# Tasks
ChEMBL 2K [ 16 ]Classification (AUC)2,35541
ToxCast [ 43 ]Classification (AUC)8,576617
Broad 6K [ 38 ]Classification (AUC)6,56732
Biogen 3K [ 12 ]Regression (MAE)3,5216
Appendix
Table 8: Summary of downstream benchmark datasets.
DatasetLearning RateDropoutPatienceScheduler
ToxCast3×10−43\times 10^{-4}0.210—
ChEMBL 2K5×10−45\times 10^{-4}0.210Cosine ( T=50T=50 )
Broad 6K5×10−55\times 10^{-5}0.120—
Biogen 3K7×10−47\times 10^{-4}0.210—
Appendix
Table 9: Best hyperparameter configurations for each downstream dataset.
More in this topic

Molecular Representation Learning

Molecular representation learning encodes chemical structures and context for prediction, generation, and discovery across chemistry and biology.

See all 207 papers

Explore similar work

CardsList

  1. Can phenotypic activity be predicted without experimental readouts?

    Oct 6, 2026Télio Cropsal, Rocío MercadoMolecular Property PredictionMolecular Representation Learning

    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

  2. Learning Molecular Representations from Cellular Phenotypes with Structure Preservation

    Aug 3, 2026Xuan Lin, Jingyu Sheng, Tengfei Ma +2Disentangled Representation LearningCross-Modal Representation Learning

    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

  3. MolEmb: Multimodal Large Language Models Can Be Strong Molecular Embedding Models

    Aug 24, 2026Xinjian Zhao, Xiangru Jian, Yaoyao Xu +4Molecular Representation LearningMultimodal Embedding

    The Chinese University of Hong Kong, Shenzhen · Shanghai Artificial Intelligence Laboratory · Cheriton School of Computer Science, University of Waterloo