TLDR: GQVis is a new, comprehensive dataset of over 2.2 million data points that pairs natural language questions about genomics data with corresponding interactive visualizations. It aims to train generative AI models to create domain-specific visualizations, featuring single queries, multi-step query chains, and rich contextual information like design justifications and alt-text, drawn from real genomics repositories.
Genomics research relies heavily on data visualization to understand complex genetic information. However, current machine learning models struggle to create these specialized visualizations because they lack proper training data. To address this, researchers have introduced GQVis, a new dataset designed to train generative AI models to create genomics data visualizations from natural language queries.
The GQVis dataset is a significant step forward, offering over 2.2 million data points. This includes 1.14 million single-query data points, 628,000 query pairs, and 589,000 multi-step query chains. Each entry in the dataset links a natural language question about genomics data with a corresponding interactive visualization, created using Gosling, a grammar-based visualization library. Beyond just the visualization, each entry also includes a data schema, alternative text for accessibility, a justification for the visualization design choices, and a figure caption. This rich context helps AI models better understand and reason about visual design.
The creation of GQVis involved a sophisticated pipeline. It started with generating abstract query templates that cover various tasks in genomic visualization, expanding on previous work like the DQVis framework. These templates use placeholders for samples, entities (like point mutation data or RNA-seq reads), and loci (specific gene locations). These placeholders are then filled with real data from genomics repositories such as 4DN, ENCODE, and Chromoscope, ensuring the queries are meaningful and grounded in actual biological data.
A unique aspect of GQVis is its inclusion of multi-step query chains. These chains, ranging from two to eight queries, simulate a sequence of analytical steps a researcher might take. For example, an initial query might ask to show data at one gene, followed by a request to compare it with data at another gene. This helps train conversational AI models to update visualizations dynamically based on follow-up user requests.
To ensure the dataset reflects the diversity of real-world user queries, the concrete queries generated from templates were paraphrased using advanced AI models like GPT-4o. This process varied the expertise and formality of the queries, enriching the dataset with a wide range of linguistic expressions for the same visualization. For instance, “What is the frequency of structural variants at FBXW7?” could be rephrased as “How common are structural variants (SVs) around FBXW7?”.
The GQVis dataset covers a broad spectrum of genomic data, including structural, functional, and epigenomic information. It supports various visualization types like point, bar, connectivity plots, heatmaps, line plots, and area plots. These visualizations can compare data across different entities, samples, and genomic locations. The dataset even incorporates complex visualizations from tools like Chromoscope, which provide interactive multiscale views of structural variation in human genomes.
Also Read:
- ConDABench: A New Benchmark for Interactive Data Analysis with Language Models
- Guiding AI Annotators: Repurposing Human Guidelines for Large Language Models
The researchers emphasize that GQVis is the first large-scale, genomics-specific dataset for natural language to visualization (NL2VIS) tasks. It provides a crucial resource and methodology for advancing generative AI-based natural language interfaces in genomics. Future work includes developing a quality assessment framework and leveraging the dataset to fine-tune large language models for genomic NL2VIS tasks. This will enable more accessible, dynamic, and interpretable genomic analysis, ultimately lowering barriers to exploratory visualization and accelerating scientific discovery. You can read the full research paper for more details. Read the full paper here.
