CellScout: Visual Analytics for Mining Biomarkers in Cell State Discovery

TL;DR

CellScout tackles the co-discovery bottleneck in cell-state biology by jointly mining associations between cell populations and biomarkers using a Mixture-of-Experts (MoE) model and a four-view visual analytics interface. The MoE mining framework optimizes discriminative power and mutual information retention under a cell-representation constraint to produce interpretable association relationships, which are explored via AI Miner, Cell Exploration, Comparison, and Verification views. Validated through expert interviews and a real-world case study, the system demonstrates the ability to reveal novel cell states and robust biomarker candidates, guiding biologists beyond traditional clustering-based approaches. The work contributes a novel MoE-based mining method for biomarker discovery and an interpretable visualization design that supports human-in-the-loop refinement, with promising directions toward multimodal knowledge integration and AI-assisted biomarker validation.

Abstract

Cell state discovery is crucial for understanding biological systems and enhancing medical outcomes. A key aspect of this process is identifying distinct biomarkers that define specific cell states. However, difficulties arise from the co-discovery process of cell states and biomarkers: biologists often use dimensionality reduction to visualize cells in a two-dimensional space. Then they usually interpret visually clustered cells as distinct states, from which they seek to identify unique biomarkers. However, this assumption is often invalid due to internal inconsistencies in a cluster, making the process trial-and-error and highly uncertain. Therefore, biologists urgently need effective tools to help uncover the hidden association relationships between different cell populations and their potential biomarkers. To address this problem, we first designed a machine-learning algorithm based on the Mixture-of-Experts (MoE) technique to identify meaningful associations between cell populations and biomarkers. We further developed a visual analytics system, CellScout, in collaboration with biologists, to help them explore and refine these association relationships to advance cell state discovery. We validated our system through expert interviews, from which we further selected a representative case to demonstrate its effectiveness in discovering new cell states.

Figures (8)

• Figure 1: This figure illustrates the overview of the problem formulation. Here, $P$ represents cell representation and can indicate the similarity among different cells based on their gene expression levels. Then, $C$ denotes individual cells, while $G$ refers to different gene types. $W$ is a matrix composed of $C$ and $G$, representing the expression levels of various genes in different cells. Based on this data and the user's domain knowledge $D$, experts aim to extract different association relationships $R$, each of which can be used to distinguish a group of cells $C_Y$ based on the uniqueness of their gene expression.
• Figure 2: The system overview of CellScout. There are three components: the Data Storage Component, the Data Analysis Component, and the Data Visualization Component.
• Figure 3: CellScout integrates MoE-based techniques to identify association relationships between cell populations and their potential biomarkers, facilitating the discovery of new cell states. There are four views that, in conjunction with the mining algorithm, help biologists understand and refine mining results based on their domain knowledge. (a) In the AI Miner View, experts can access results generated by MoE-based techniques. We showcase various mined association relationships and highlight the cell populations that might be most relevant to them. Additionally, we display the importance of each gene within those relationships. (b) The Cell Exploration View helps experts effectively explore single-cell gene expression data based on their two-dimensional representations. (c) In the Comparison View, experts can compare different cell populations and assess the relevance of each association relationship. (d) In the Verification View, experts can perform statistical validation.
• Figure 4: The gene importance scores computed by our model for each association relationship can be visualized by clicking on the corresponding association node in the AI Miner View or Cell Exploration View.
• Figure 5: The alternative design represents gene importance within each mined association relationship using a matrix visualization.