Title: Facilitating the Creation of Interactive Multiview Genomics Visualizations Through Agentic Authoring

Abstract:

The heterogeneity of genomics datasets, combined with the specific nature of scientific inquiries and analytical objectives, frequently necessitates the use of highly specialized visual representations. Consequently, researchers are often compelled to either customize existing tools or develop new visualizations from scratch to suit their specific data. However, current solutions present significant barriers: many are restricted in their customization capabilities, while others demand extensive programming knowledge or a steep learning curve. Even the most flexible tools presuppose a level of visualization expertise that many users do not possess.

Although natural-language conversational interfaces represent a promising avenue for democratizing the creation of complex visualizations, applying agentic and large language model (LLM) approaches to this domain introduces unique hurdles. Specifically, genomics visualizations often involve the integration of diverse data types and the coordination of multiple linked, interactive views. These complexities suggest a need for more structured, LLM-based frameworks.

This study first delineates the boundaries of vanilla LLM capabilities in genomics visualization, identifying eight distinct dimensions of quality where such models succeed or fail. We subsequently evaluate six distinct generation schemes—ranging from direct generation and a fixed pipeline to four agentic configurations that vary by the number of specialist agents and the inclusion of a reviewer role. Our evaluation covers 159 case studies, encompassing three tiers of query ambiguity and specification complexity. All methods utilize the Gosling visualization grammar to ensure structured output.

Our findings indicate that agentic iteration significantly enhances perceived quality compared to baseline approaches. However, increasing the complexity of the agent architecture does not yield further improvements. We conclude by discussing the implications of these results for the design of agentic systems tailored to domain-specific visualization authoring. Supplementary materials can be accessed at https://osf.io/uqe83.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC