CP-Agent Turns Drug Cell Images Into Human-Readable Mechanism Reports
CP-Agent pairs a CLIP-style image-metadata encoder with an agentic LLM to explain why a drug changed cell shape -- hitting F1 of 0.896 on mechanism-of-action classification
Takeaways
- CP-Agent is a new agentic multimodal LLM that generates human-readable mechanism-of-action reports from Cell Painting microscopy images. [Paper]
- Its core module, CP-CLIP, jointly embeds fluorescence images and experimental metadata (cell line, dose, schedule), reaching F1 of 0.896 on MoA classification.
- Prior Cell Painting workflows ignored experimental context like cell line and dosing, limiting generalization -- CP-Agent bakes that context directly into the embedding.
- An agentic layer with FeatRank and ReportGen sub-agents converts embeddings into structured reports linking morphological changes to biological mechanisms.
- Evaluated on the JUMP-CP dataset (116,000+ compounds) across multiple cell lines; also tests a fine-tuned VISTA segmentation backbone for upstream image quality.
- No public code or model weights released yet; hallucination risk from external knowledge retrieval is a noted limitation.
CP-Agent is a new agentic multimodal system that can look at fluorescence microscopy images of drug-treated cells, read the experimental context, and write a structured, human-readable report explaining why the cells changed shape. That last part -- the "why" -- is what makes it different from every prior tool in this space.
The Assay That Became a Bottleneck
Cell Painting is a high-content, multiplexed fluorescence imaging assay designed to capture the morphological state of cells across multiple biological compartments simultaneously. The protocol multiplexes six fluorescent dyes, imaged in five channels, to reveal eight broadly relevant cellular components or organelles. The result is a rich visual fingerprint of what a drug does to a cell.
High-throughput morphological profiling of small molecule libraries using Cell Painting has received increasing attention in drug discovery research. Compared to conventional high-throughput screening of a single biological target, morphological profiling with high content imaging offers the advantage of identifying multiple biological activities of small chemical compounds simultaneously, promising to substantially accelerate the early stage of the drug discovery process.
Cell Painting has become an assay of choice for exploring mechanism of action (MoA) and target identification studies in drug library screens, especially valuable when the target is complex, considered undruggable, or multipart. The problem is that the existing analysis pipeline is a black box: you get a cluster of compounds that look similar, but no explanation of the biology driving that similarity.
What the Field Was Missing
Cell Painting combines multiplexed fluorescent staining, high-content imaging, and quantitative analysis to generate high-dimensional phenotypic readouts to support diverse downstream tasks such as MoA inference, toxicity prediction, and construction of drug-disease atlases. However, existing workflows are slow, costly, and difficult to interpret. Approaches for drug screening modeling predominantly focus on molecular representation learning, while neglecting actual experimental context -- such as cell line or dosing schedule -- limiting generalization and MoA resolution.