Dynamic AI Agent-Built Visualization Tools for Scientific Inquiry
Year: 2026
Type: article
Abstract: Background Physician-scientists receive extensive education to conduct leading-edge research in their domain. Modern, data-intensive science demands a third competence in software alongside PhD-level research experience and MD-level clinical acumen. In practice, analysis depth may be limited by the scientist's own technical skills or by lack of access to specialist collaborators. Scientific questions that are easy to articulate but time-intensive to pursue go unanswered. Even capable specialists often cannot justify writing bespoke software for a question that will be asked only once. Approach Scientists can prompt AI agents to build custom data visualizations during analysis, allowing them to iteratively refine and tailor analytical approaches in near real-time. The scientist remains responsible for hypothesis generation and interpretation, while AI agents build the visualizations. This approach allows rapid exploration of the analysis space, whether through deeper interrogation of a single method or implementation of a wider variety of them. Results We built several variations of a visualization tool to inspect the output of a single hierarchical clustering approach applied to high-dimensional spectral flow cytometry data. The resulting views revealed that the cell population of interest was being blurred with phenotypically distinct cells, a finding that had not been apparent from summary statistics alone. The finding prompted us to broaden the set of clustering algorithms used for analysis. Broader Impact Tool-building and data visualization have always been part of the scientific process, but traditionally, investment in customized tools had to be justified by recurring needs. AI agents have changed the calculus of when a tool is worth building. We argue that scientists should treat tool-building as a dynamic part of data analysis rather than relying upon a static set of processes and approaches. In doing so, researchers gain access to analysis approaches that were previously out of reach.
Authors: Corin, Dan, Salzman, Gabriel, Meyer, Everett
Institutions: Stanford University, University of Colorado Anschutz
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7213525009