ClusterChirp: A GPU-accelerated Web Server for Natural Language-Guided Interactive Visualization and Analysis of Large Omics Data

Tabular datasets are commonly visualized as heatmaps, where data values are represented as color intensities in a matrix to reveal patterns and correlations. However, modern omics technologies increasingly generate matrices so large that existing visual exploration tools require downsampling or filtering, risking loss of biologically important patterns. Additional barriers arise from tools that require command-line expertise, or fragmented workflows for downstream biological interpretation. We present ClusterChirp, a web-based platform for real-time, interactive exploration of large-scale data matrices enabled by GPU-accelerated rendering and parallelized hierarchical clustering using multiple CPU cores. Built on deck.gl and multi-threaded clustering algorithms, ClusterChirp supports on-the-fly clustering, multi-metric sorting, feature search, and adjustable visualization parameters for interactive explorations. Uniquely, a natural language interface powered by a Large Language Model helps users perform complex operations and build reproducible workflows from conversational commands. Furthermore, users can select clusters to explore interactive within-cluster correlation networks in 2D or 3D, or perform functional enrichment through biological knowledge bases. Developed with iterative user feedback and adhering to FAIR4S principles, ClusterChirp empowers researchers to extract insights from high-dimensional omics data with unprecedented ease and speed. This website is freely available at clusterchirp.mssm.edu, with no login required.

💡 Research Summary

ClusterChirp is a web‑based platform that enables real‑time, interactive exploration of massive omics matrices without the need for down‑sampling. The system combines GPU‑accelerated heatmap rendering (via WebGL and the deck.gl library) with a parallelized hierarchical clustering engine that distributes O(n²) distance calculations across multiple CPU cores using joblib and Numba JIT compilation. This architecture allows datasets containing tens of thousands of features and up to millions of cells to be visualized at 60 fps, while clustering operations that previously required hours are completed in minutes.

The front‑end, built with React and TypeScript, provides a rich set of interactions: pan, zoom, tooltips, dynamic search, metadata‑aware filtering, and adjustable visual parameters. A standout feature is the natural‑language chatbot powered by OpenAI’s GPT‑4o‑mini model. Users can issue conversational commands such as “cluster genes using Pearson correlation and show the top 20 clusters”; the LLM parses the request, incorporates the current session state (filters, sorting, clustering settings), and returns a standardized JSON action object. The back‑end validates this object and executes the corresponding front‑end or server‑side operation. In benchmark tests on a 77‑sample plasma‑protein dataset, the chatbot achieved a 95.66 % success rate across 45 representative commands, with average response times under two seconds. A rule‑based fallback parser ensures functionality when the LLM service is unavailable.

For deeper biological insight, ClusterChirp offers on‑the‑fly construction of intra‑cluster correlation networks. After removing features with >20 % missing values and retaining the most variable 75 % of the remaining features, pairwise Pearson correlations are computed in parallel. Users can visualize the resulting network in 2D (Sigma.js with ForceAtlas2 layout) or 3D (Three.js with Leiden community detection). Large networks (>1,000 nodes) are processed in Web Workers to keep the UI responsive.

Functional enrichment is seamlessly integrated via the Enrichr API. When a cluster is selected, gene identifiers are harmonized to HUGO symbols (or mapped from Olink protein panels) and submitted to Enrichr; results appear in a new browser tab, providing immediate pathway and disease‑association context.

ClusterChirp adheres to the FAIR4S principles: the source code for both front‑end and back‑end is publicly available on GitHub under an APGL‑3.0 license; the web service is freely accessible without registration; standard tab‑delimited formats are supported for input and output; and comprehensive documentation, tutorials, and example datasets are provided to promote reproducibility. Data handling follows HIPAA‑compliant practices on Mount Sinai’s Minerva server, with session‑specific temporary storage and automatic cleanup.

User‑centered design guided development: ten domain experts participated in iterative interviews, low‑fidelity prototypes, and high‑fidelity usability testing, leading to refinements in UI layout, natural‑language interaction, and network visualization features. Subsequent evaluation with five additional researchers confirmed the platform’s usability and performance.

In summary, ClusterChirp bridges the gap between high‑throughput omics data generation and biological interpretation by delivering GPU‑driven, scalable visual analytics combined with an AI‑powered conversational interface. It eliminates the need for data down‑sampling, reduces clustering latency from hours to minutes, and integrates downstream functional analysis within a single, accessible web environment, thereby accelerating discovery and democratizing large‑scale omics exploration.