Can LLMs Bridge Domain and Visualization? A Case Study on High-Dimension Data Visualization in Single-Cell Transcriptomics

• DOI: 10.31219/osf.io/qtsak_v2
• Publication Date: 2025-04-02T20:55:24Z
• Authors: Qianwen Wang; Xinyi liu; Nils Gehlenborg
• Keywords: Bridge (graph theory)

While many visualizations are built for domain users (e.g., biologists, machine learning developers), understanding how used in the has long been a challenging task. Previous research relied on either interviewing limited number of or reviewing relevant application papers visualization community, neither which provides comprehensive insight into ``in wild'' specific domain. This paper aims to fill this gap by examining potential using Large Language Models (LLM) analyze usage literature. We use high-dimension (HD) data sing-cell transcriptomics as test case, analyzing 1,203 that describe 2,056 HD with highly specialized terminologies biomarkers, cell lineage). To facilitate analysis, we introduce human-in-the-loop LLM workflow can effectively large collection and translate domain-specific terminology standardized task abstractions. Instead relying solely LLMs end-to-end our enhances analytical quality through 1) integrating image processing traditional NLP methods prepare well-structured inputs three targeted subtasks (\ie, translating terminology, summarizing analysis tasks, performing categorization), 2) establishing checkpoints human involvement validation throughout process.The results, validated expert interviews set, revealed often overlooked aspects visualization: trajectories spaces, inter-cluster relationships, dimension clustering.This stepping stone future studies seeking bridge between design usage.