AI Deciphers How Biologists Visualise Complex Data

Designing software for scientists is challenging because understanding how they use visualisations "in the wild" is surprisingly difficult. Traditional research relies on small-scale interviews, which fail to capture the breadth of real-world usage. To bridge this gap, researchers have developed a "human-in-the-loop" workflow utilising Large Language Models (LLMs).

The team applied this method to single-cell transcriptomics, a field laden with complex data. By integrating image processing with human validation checkpoints, the AI analysed 1,203 papers containing 2,056 high-dimensional visualisations. Crucially, the system translated dense terminology—such as "cell lineage" and "biomarkers"—into standardised data abstractions that visualisation experts can understand.

The results were illuminating. Validated by expert interviews, the analysis revealed that designers often overlook key requirements, specifically regarding trajectories in high-dimensional spaces and inter-cluster relationships. This research demonstrates that LLMs can effectively translate between specialised domain needs and general visualisation design, ensuring future tools are better suited to scientific discovery.