Integrating Segmented Cell Imaging and Molecular Networks for Drug-Specific Analysis in CM4AI

Agents4Science 2025 Conference Submission321 Authors

Keywords: Multimodal learning, Protein–protein interaction networks, Cell morphology, Network biology, Deep learning

TL;DR: We present an AI-assisted pipeline that integrates single-cell microscopy embeddings with protein–protein interaction networks to reveal drug-specific molecular reprogramming with enhanced resolution, reproducibility, and translational relevance.

Abstract: Linking cellular morphology to molecular interaction networks remains a central challenge in biomedical Artificial Intelligent (AI). We present a cell-centric framework that integrates object-level detection with Vision Transformer (ViT) embeddings from microscopy with protein--protein interaction (PPI) representations to construct biologically interpretable hierarchies and reveal condition-specific network reconfiguration. Using a semi-automated, agent-oriented workflow, segmentation is executed via an interactive Large Language Model (LLM)-driven agent bridged to high-performance computing, while embedding, integration, and hierarchy construction proceed through reproducible human--LLM collaboration with auditable prompts, code generation, and logged execution. Applied to 12853 high-content images spanning Untreated, Vorinostat-, and Paclitaxel-treated conditions, the approach preserves global biological structure while sharpening signal fidelity relative to whole-image baselines, enabling single-cell resolution of heterogeneity. Across all conditions, the modified pipeline maintained >95\% concordance with baseline hierarchies. Gene Ontology analyses recover drug-consistent pathways (e.g., chromatin regulation for Vorinostat; microtubule-associated processes for Paclitaxel) and yield more selective enrichment profiles. The framework establishes a scalable foundation for multimodal integration with additional omics layers and for prospective validation of predicted network rewiring in precision medicine contexts.

Submission Number: 321