Abstract
Exploratory data visualization is a routine requirement in bioinformatics research, yetproducing and customising domain-specific charts such as volcano plots typically re
quires scripting proficiency in R or Python. This creates a dependency on a small
number of visualization-skilled individuals within research teams, slowing the iterative
sensemaking cycle for newcomers and domain experts alike. Recent advances in large
language models (LLMs) suggest that conversational interfaces can lower this barrier by
accepting analytical requests in natural language and translating them into executable
data operations and chart specifications. However, deploying such systems in a real
bioinformatics context introduces challenges around ambiguity handling, hallucination
risk, transparency of reasoning, and data privacy that purely technical approaches do
not fully address. This thesis presents Viz-LLM-Bot, a conversational visualization assis
tant that connects an LLM to a tabular differential gene expression (DEG) dataset and
enables bioinformaticians to request and iteratively refine visualizations through natu
ral language chat. The system was designed and implemented as a customer-requested
prototype at Datavisyn GmbH and evaluated with bioinformatician participants from
within the organisation. The design process was grounded in the human-AI interac
tion and visualization literature, with four core design constraints (domain grounding,
ambiguity clarification before generation, visible reasoning, and iterative conversational
state) translated into concrete architectural and interaction decisions. The prototype
presents a three-panel interface in which a DEG results table, a visualization panel,
and a chat interface are simultaneously visible. Users can filter, sort, group, and visu
alize the dataset entirely through natural language commands. The system exposes a
decision trace alongside each output, showing the sequence of reasoning steps and data
operations that produced the result. When a request is underspecified, the system issues
a targeted clarification question rather than committing silently to a default interpre
tation. The user study was conducted with eight bioinformaticians at Datavisyn GmbH
using a within-company evaluation design built around realistic, domain-relevant ana
lytical tasks. Participants worked through a set of defined tasks using the live prototype
in a think-aloud session, after which task completion was recorded alongside perceived
usability assessed with the System Usability Scale (SUS), perceived trust assessed with
the Trust between People and Automation (TPA) questionnaire, and targeted Likert
items covering perceived accuracy and transparency of the system’s outputs. The find
ings suggest that LLM-driven conversational visualization assistance is both technically
feasible and, in terms of perceived usability and trust, broadly positively received by
bioinformatician users in an industrial setting, while also revealing specific interaction
design challenges around prompt formulation, output verification, and the boundaries of the system’s domain knowledge. The thesis concludes with a discussion of the conditions
under which such systems can be responsibly integrated into bioinformatics workflows
and identifies the most productive directions for future development.
| Datum der Bewilligung | 2026 |
|---|---|
| Originalsprache | Englisch (Amerika) |
| Gradverleihende Hochschule |
|
| Betreuer/-in | Mandy Keck (Betreuer*in) |
Studiengang
- Interactive Media
Zitieren
- Standard