Title: CP-Agent: Enabling Context-Aware Multimodal Reasoning for Cellular Morphological Profiling Amidst Chemical Perturbations

Abstract

The Cell Painting technique integrates high-content imaging, multiplexed fluorescent staining, and quantitative analysis to produce high-dimensional phenotypic data. These data support a variety of downstream applications, including the construction of drug-disease atlases, toxicity prediction, and mechanism-of-action (MoA) inference. Despite its utility, current workflows are often hampered by high costs, slow processing speeds, and a lack of interpretability. Furthermore, most existing drug screening models prioritize molecular representation learning, frequently overlooking critical experimental contexts such as dosing schedules and specific cell lines. This oversight restricts both the generalizability of the models and the resolution of MoA identification.

To address these challenges, we present CP-Agent, an agentic multimodal large language model (MLLM) designed to generate human-interpretable rationales for morphological changes in cells resulting from drug perturbations. The foundation of CP-Agent is CP-CLIP, a context-aware alignment module that jointly embeds experimental metadata with high-content images. This approach facilitates robust discrimination of treatments and MoAs, reaching a peak F1-score of 0.896. By combining CP-CLIP outputs with agentic reasoning and tool usage, CP-Agent synthesizes these insights into structured reports. These reports are intended to refine hypotheses and guide experimental design, demonstrating the model’s potential to accelerate drug discovery through scalable, interpretable, and context-aware phenotypic screening that streamlines iterative hypothesis generation.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC