Integrating chemical structures as treatments improves representations of microscopy images for morphological profiling

Technical Summary: MICON – Molecular-Image Contrastive Learning for Morphological Profiling

Problem Statement
High-throughput microscopy screens, such as Cell Painting, are critical for drug discovery, enabling the quantification of cellular morphological changes in response to chemical or genetic perturbations. This process, known as morphological profiling, converts images into feature vectors (representations) to compare cell phenotypes. While deep learning has improved upon traditional hand-crafted features (e.g., CellProfiler), most existing self-supervised methods learn exclusively from image data. This approach ignores the inherent multimodal nature of these experiments, where a specific chemical compound (perturbation) induces the observed phenotypic change. Furthermore, microscopy data is plagued by "batch effects"—technical variations arising from different instruments, days, or sources—that can confound biological signals. Current multimodal approaches that attempt to align images and compounds directly (e.g., via CLIP objectives) often struggle because phenotypic responses to the same compound vary by cell type, creating a tension where a single compound representation must align with multiple disparate phenotypic representations.

Methodology: MICON Framework
The authors introduce MICON (Molecular-Image Contrastive Learning), a self-supervised representation learning framework that explicitly models chemical compounds as treatments that induce transformations on cell images, rather than treating images and compounds as directly comparable entities in a shared space.

The framework consists of three primary components:

1. Encoders:
   • An Image Encoder ($h_{img}$), based on a ResNet101 pre-trained on ImageNet, processes microscopy images.
   • A Compound Encoder ($h_{comp}$), a multi-layer perceptron (MLP), processes chemical compounds represented as ECFP4 fingerprints.
2. Fusion Module: A two-layer MLP ($F$) that takes the concatenation of a negative control image representation and a compound representation to predict the representation of the image as if it had been treated with that compound.
3. Loss Functions: The model is trained using two contrastive objectives:
   • Perturbation-Aware Contrastive Learning (PaCLR) on Real Images: Unlike standard contrastive learning (e.g., SimCLR) which treats all other images in a batch as negatives, PaCLR uses metadata to define positive pairs. Images treated with the same perturbation (even from different batches/sources) are positive pairs. Crucially, negative control images (DMSO) from the same microscopy batch as the perturbed images are used as negative pairs. This forces the model to learn biological signals invariant to batch effects while distinguishing perturbations from technical noise.
   • Predicted PaCLR Loss: This objective aligns the predicted representations (generated by the fusion module) with the real representations of cells actually treated with the compound. This loss specifically integrates chemical structure information, training the model to simulate phenotypic transformations.

The final loss is an equal weighting of the real PaCLR loss and the predicted PaCLR loss.

Key Contributions

• Novel Multimodal Formulation: MICON proposes modeling compounds as treatments that transform image representations, avoiding the pitfalls of direct image-compound alignment where phenotypic variability creates alignment conflicts.
• Robustness to Batch Effects: By explicitly incorporating negative controls from the same microscopy batches into the contrastive learning sampling strategy, MICON learns representations robust to technical variation across sources and batches.
• Ablation of Multimodal Integration: The study isolates the contribution of chemical structure information by comparing the full MICON model against a version trained only with the PaCLR loss (image-only) and against baselines like SimCLR, CLIP, and DINOv3.

Results
The authors evaluated MICON on the JUMP Cell Painting Consortium (cpg0016) dataset, utilizing rigorous data stratification to test generalization to unseen sources (Out-of-Distribution, OOD) and unseen compounds.

• Superior Performance on Reproducible Effects: In compound-replicate matching tasks (identifying if two images share the same treatment), MICON significantly outperformed CellProfiler, SimCLR, CLIP, and DINOv3. On the Positive Control (POS-CTL) in-distribution dataset, MICON achieved 62.24% accuracy (1-NN, Not-Same-Batch) compared to 36.56% for CellProfiler and 37.77% for SimCLR.
• Generalization to Unseen Sources: When tested on a source completely unseen during training, MICON maintained significant performance gains over baselines (51.57% vs. 39.26% for CellProfiler), whereas other self-supervised methods (SimCLR, CLIP, DINOv3) dropped to or below CellProfiler performance, suggesting they overfit to training sources.
• Generalization to Unseen Compounds: In OOD settings involving 184 unseen compounds, MICON again outperformed all baselines (4.92% vs. 3.53% for CellProfiler), demonstrating that integrating chemical information improves generalization to new perturbations.
• Robustness to Batch Correction: Even after applying standard post-hoc batch correction (spherizing), MICON continued to outperform other self-supervised methods, indicating its performance stems from superior phenotypic recognition rather than just batch effect mitigation.
• Mechanism of Action (MoA) Enrichment: MICON showed the highest enrichment for compounds sharing the same MoA, suggesting better clustering of biologically similar phenotypes.
• Synthetic Prediction: The model demonstrated the ability to predict phenotypic effects of unseen compounds using generated representations, though performance on unseen compounds was modest compared to seen compounds.

Significance and Claims
The paper claims that MICON establishes a new direction for representation learning in morphological profiling by explicitly accounting for the multimodal nature of screening data. The authors argue that:

1. Multimodal Integration is Essential: Incorporating chemical structure information provides small but consistent, statistically significant improvements over image-only learning.
2. Modeling Strategy Matters: Treating compounds as "treatments" that transform images (MICON) is superior to directly aligning images and compounds in a single space (CLIP-style approaches).
3. Robustness to Real-World Variability: MICON's performance gains are not merely due to better batch correction but reflect a more robust capture of biological phenotypic signals, particularly in scenarios involving unseen data sources or compounds.

The authors remain modest regarding the predictive capability of the model, noting that while it can synthesize phenotypic effects, the current performance on unseen compounds is low, and further optimization of the loss weighting and hyperparameters is required before such predictive capabilities can be practically deployed for triaging compounds.