BioVis@ISMB 2026 Program

Cell population plots are used to visualize cell types, states or clusters, and compare cell types within and between samples across conditions. Cell populations are traditionally shown using a stacked bar chart approach, with samples as bars and cell types as colored segments with lengths corresponding to the number or proportion of cells. These visualizations do not scale well with increasing numbers of samples and cell types, making it hard to identify and compare cell types. This is becoming a larger issue as single-cell atlas studies combine more and larger samples with more and rarer cell types.

We conducted a design study to evaluate the user tasks and need for cell population plots, including within-sample and between-sample comparisons and metadata alignment. This highlighted color interpretability limitations and a need for interactive filtering, sorting, and grouping.

Here we introduce scellop, a redesigned flexible cell population viewer (https://github.com/hms-dbmi/scellop). scellop combines a heatmap of cell counts with expandable embedded bar charts to support both global pattern detection and detailed investigation of samples.

scellop is available as a Python and JavaScript package and integrates in Jupyter environments and web applications. It supports common single-cell data formats and allows flexible configuration and exports. Together, it enables scalable exploration of single-cell datasets and improves comparisons compared to traditional stacked bar chart approaches.

The National Center for Biotechnology Information (NCBI) provides a suite of interactive, web-based tools to support visualization and analyses of genomic data across a wide range of organisms. Together, the Genome Data Viewer (GDV), Comparative Genome Viewer (CGV), and Multiple Comparative Genome Viewer (MCGV) offer complementary approaches to exploring data at a single genome level to pairwise comparisons to multi-genome comparisons.

GDV (https://ncbi.nlm.nih.gov/gdv) serves as NCBI's flagship genome browser, displaying gene annotations from multiple sources such as RefSeq, GenBank and Ensembl, variation data from NCBI and EVA, RNA-seq expression, and user-provided custom tracks for over 4500 eukaryotic genome assemblies. GDV integrates with other NCBI resources including BLAST, dbGaP, GEO, and ClinVar, enabling comprehensive genomic analyses within a single platform.

For comparative genomics, CGV (https://ncbi.nlm.nih.gov/cgv) visualizes pairwise whole-genome assembly alignments using an interactive ideogram and dotplot views, enabling exploration of structural differences such as inversions and translocations, as well as examining synteny of homologous genes when structural annotations are available. CGV currently supports over 1,600 pairwise alignments across more than 750 species, with new alignments added continuously in response to requests from the scientific community.

MCGV (https://ncbi.nlm.nih.gov/mcgv) extends comparative visualization to multiple genomes simultaneously, displaying synteny blocks and sequence conservation relative to an anchor assembly, with support for pangenome datasets. Developed as part of the NIH Comparative Genomics Resource (CGR) initiative, MCGV—together with GDV and CGV—forms an interconnected ecosystem enabling researchers to navigate genomic data across biological scales without the need for local software installation or data preprocessing.

The Systems Biology Graphical Notation (SBGN) provides standardized visual languages for representing complex biological processes, facilitating communication, reproducibility, and model sharing in systems biology. However, generating high-quality SBGN maps from scratch can be challenging, particularly for new users, due to the learning curve associated with SBGN editors and the difficulty of translating informal ideas into structured diagrams. To address this, we present SBGNFlow, an AI-assisted and interactive workflow that streamlines SBGN map generation and refinement through three key steps, supporting both Process Description (PD) and Activity Flow (AF) languages. The first step enables automatic conversion of hand-drawn SBGN sketches into SBGN-ML format using large language models with in-context learning. Quick correction of minor recognition errors and text-based editing are supported through an interactive interface, while biological identifiers are mapped automatically to facilitate annotation. Second, SBGNFlow supports flexible merging and splitting of maps. Digitized maps can be merged with existing ones to create larger networks in incremental steps by identifying shared nodes and edges, or reorganized into smaller components as needed, while preserving the user’s mental map. Finally, we introduce layout refinement methods. A user-guided layout algorithm allows sketch-based hints to influence the arrangement of the entire network or selected subgraphs, while a polishing step improves readability by aligning edges orthogonally or diagonally and organizing nodes by functional role (input, output, modifier). Together, these features provide an end-to-end solution for transforming informal sketches into structured, publication-ready SBGN maps, lowering the entry barrier for new users while offering flexible control for experts.

InterSCellar (https://pypi.org/project/InterSCellar/) is an open-source Python package for surface-based cell-cell interaction analysis in 3D spatial omics data. Current practices for estimating cell-cell interactions are simply centroid-based and ignore the shape irregularities between different cell types. To address this, InterSCellar implements two core functionalities: (1) construction of cell-neighbor graphs through surface-based detection of adjacent cell pairs; (2) computation of intercellular spaces between neighbors as physical volumes. Therefore, InterSCellar more accurately identifies subcellular interaction contexts grounded in spatial adjacency, redefining how cell-cell interactions are detected, quantified, and interpreted.

Applied to highly-multiplexed (>50-channels) 3D CyCIF melanoma datasets of clinical melanoma samples, InterSCellar reveals interaction-specific structural and molecular patterns across different stages of disease progression. We identify distinct shifts in neighborhood organization, with invasive tumors forming more self-associated or immune-enriched clusters, while in situ tumors retain stronger epithelial and stromal architecture. Proteomic expression within interaction volumes uncovers enrichment patterns specific to neighboring cell profiles and differential localization of structural markers across tumor–epithelial interfaces. These results enable downstream analyses such as network-based characterizations of tissue microenvironments and identification of spatial niches associated with disease states.

InterSCellar (https://github.com/hms-dbmi/InterSCellar) is interoperable with standard data structures including OME-NGFF and scverse’s SpatialData, and integrates with visualization tools such as Vitessce and Napari. By transforming high-dimensional 3D images into surface-resolved interaction graphs and volumes, InterSCellar bridges the analytical gap between segmentation and biological insight.

Spatial transcriptomics enables the study of gene expression while preserving spatial context within tissue, generating datasets that are high-dimensional, multimodal, and spatially structured. These characteristics introduce challenges for visualization, requiring methods that support reasoning across molecular, cellular, and tissue scales during exploration, validation, and communication of biological hypotheses. Despite a growing ecosystem of tools, there remains limited understanding of how visualizations encode data, support biological inquiry, and scale with analytical complexity. To address this gap, we developed a survey of visualization practices in spatial transcriptomics, grounded in the systematic coding of over 3,000 figure panels across 170 published analysis and visualization tools. After excluding benchmarking and schematic ones, over 1,800 were retained for analysis. We systematically coded each figure using a taxonomy inspired by Munzner’s what–why–how framework, capturing three dimensions: data representation (what), analytical tasks (why), and visualization design (how), with interaction characterized as an additional dimension. Through this analysis, we identify recurring design patterns; Visualization is predominantly oriented toward questions of tissue structure, followed by tasks focused on identifying cell types. Comparative analysis is most commonly supported through juxtaposition using small multiples, typically to contrast gene expression patterns within tissue. We also observe that some practices mirror those used in single-cell transcriptomics, discarding spatial context. By linking biological research questions to visualization design, this review establishes a task-oriented perspective on visualization practice in spatial transcriptomics. Our findings highlight design biases and missed opportunities, providing guidance for developing tools that better support spatial reasoning and complex biological inquiry.

Modern sequencing technologies generate millions of antibody sequences per experiment, generating complex datasets. Current approaches usually present mutations as static tables or scatter plots, failing to preserve sequence context or integrate predictions from multiple protein language models. Consequently, scientists must manually piece together insights, introducing inefficiency and risking losing critical insights. We present an interactive visualization system addressing fundamental challenges of multi-scale navigation, multi-objective optimization, and exploratory analysis of mutation spaces. Our web-based platform integrates three coordinated views, an interactive sequence navigator with Complementarity Determining Regions (CDR) highlighting, an adaptive multi-position heatmap, and a selection drawer for detailed inspection. This enables intuitive exploration from overview to mutation-level detail. Beyond antibody engineering, this work demonstrates generalizable visualization strategies for biological sequence optimization problems involving multi-model predictions and multi-objective decision-making. The system exemplifies how thoughtful visualization design can transform complex bioinformatics predictions into actionable insights while maintaining critical spatial context throughout exploratory analysis, using simulated antibody sequence data.

Clustering is a fundamental step in many spatial transcriptomics (ST) analysis pipelines, including Space Ranger. Clusters are often used for manual annotations, biological interpretations, and downstream analyses such as differential gene expression analysis (DGEA). However, most clustering methods are stochastic and can yield different partitions even when input and parameter are fixed except for a random seed. The key question is whether clustering are stable or merely artifacts of a particular run? In this talk, we present a model-agnostic framework for quantifying and visualizing clustering instability in ST analyses. We apply this framework to 18 publicly available ST benchmark samples with ground truth labels and 7 clustering methods. Our results revealed that unstable clusters were pervasive across benchmarks, and that cluster stability depended more on the sample itself than the clustering method. Unstable clusters were typically associated with lower purity, lower gene expression coherence, and lower DGEA consistency; while stable clusters showed the reverse. A central contribution of our framework is the visualization of clustering instability in spatial context. We introduce spatial instability maps that allow users to inspect unstable tissue regions, compare instability patterns across methods and samples, and relate instability to biological annotations and downstream analyses. These visualizations help biologists identify regions where cluster-based interpretation should be treated cautiously or where alternative analysis strategies may be needed. Ongoing work extends this framework into an interactive tool for exploring unstable regions, inspecting cluster behavior across repeated runs, and using stability information to guide study conclusions.

Genomics visualizations require the integration of heterogeneous data types, coordinated interactive views, and domain-specific constraints, making authoring more difficult than standard chart generation. Although many visualization tools are available, they are typically either limited in customization or require extensive learning or programming effort, and even when a tool is sufficiently expressive, users may lack the visualization expertise to produce effective designs. Large language models are increasingly used for automatic visualization generation from natural language, but remain limited when applied to complex, domain-specific data such as genomics, where their complexity might also make it more difficult for users to articulate their needs precisely in natural language.
We investigate how agent-based LLM schemes can support complex genomics visualization authoring in two steps. We first characterize where vanilla LLM generation succeeds and fails using Gosling, a declarative grammar for genomics visualization, across 159 cases spanning three complexity levels and three query scenarios. We identified eight quality dimensions covering encoding, layout, interaction, and presentation, and found that persistent failure modes emerge as complexity increases, motivating more structured authoring approaches with iteration and correction.
We then compare six authoring schemes ranging from direct generation and a fixed pipeline to four agentic configurations, varying in the number of specialist agents and the presence of a reviewer. Our results show that agentic schemes substantially outperform both baselines on perceived quality, while more complex agent architectures did not improve over a single iterative agent, which achieved comparable quality at the lowest cost. Our findings suggest that specialized grammars and flexible agents are complementary: a grammar acts as a guardrail that constrains the solution space, and the agent flexibly searches within those limits. Realizing this potential will require finer-grained tools such as per-track validation and interaction-level debugging, allowing agents to inspect and correct more precisely than static visual and spec-level feedback currently supports.

Network models offer a flexible and effective framework for analyzing large-scale biomedical data. Interactive visualization plays a central role in this process, with tools such as Cytoscape, Gephi, and VRNetzer widely used to explore and interpret network biology models. However, the complexity of large biomedical networks presents a significant challenge for developing intuitive visualizations that enable rapid identification of patterns within intricate topologies. Existing tools, while powerful, do not always fully address the interactivity, scalability, and integrative analysis required for increasingly large and heterogeneous datasets, which often demand customized visual configurations and workflows.
Here, we present WWizNet, a fully interactive web-based application that integrates scalable visualization with advanced analytical functionality in a unified environment. The platform supports real-time rendering of large networks alongside rich metadata annotation, including dimensionality reduction approaches such as UMAP applied to functional annotations (e.g., Gene Ontology or Human Phenotype Ontology), enabling discovery of relationships beyond topology. Classical network analysis methods, including community detection and random walk with restart, are implemented to explore structural organization and prioritize relevant nodes. Additionally, we introduce a plugin for visualization of AlphaFold protein structures, with emphasis on amino acid highlighting for phosphoproteomic data and disease-associated variants.
The application further incorporates seemingly integrated metadata analysis tools, including interactive heatmaps and UpSet plots that respond dynamically within a unified node selection environment. This unified design enables iterative, exploratory workflows, facilitating efficient hypothesis generation, validation, and knowledge discovery within a single, scalable platform.

Chromatin accessibility profiling across hundreds of tumor samples[1] and millions of candidate regulatory elements[2] reveals regulatory programs that distinguish cancer subtypes, yet most analyses reduce this complexity to static figures that cannot be interrogated further. Purpose-built data portals support predefined queries but do not empower users to reshape analyses on the fly. We present Epifolio[3,4], a suite of reactive Marimo[5] notebooks that couple curated multi-modal views of the TCGA regulatory landscape with the full analytical flexibility of a computational notebook. Marimo's reactive execution model, combined with custom visualization widgets built on Anywidget[6,7], lets users transition between a purpose-driven "app" interface and an editable notebook interface for bespoke analyses. Since all processed data artifacts are hosted statically, Epifolio runs across a range of environments with zero configuration. Multiple notebooks guide exploration across scales: cohort-level dimensionality reduction (NMF, PCA) linked to clinical metadata, pairwise accessibility comparisons between patients, and side-by-side inspection of healthy and tumor signal tracks at specific loci. The dataflow model propagates selections across linked views, so users can lasso a cluster in a UMAP embedding, immediately see corresponding NMF loadings, clinical annotations, and underlying signal, then modify code to pursue new questions without leaving the notebook. To validate this workflow, we recovered the PAM50 basal versus non-basal breast cancer distinction -traditionally derived from gene expressions- from unsupervised decomposition of chromatin accessibility alone. All notebooks are downloadable, self-contained, and executable. Epifolio's design demonstrates that reactive notebooks can bridge the gap between curated data resources and open-ended multi-omic exploratory analysis.



References:
1. M. Ryan Corces et al. ,The chromatin accessibility landscape of primary human cancers.Science362,eaav1898(2018). DOI:10.1126/science.aav1898
2. Moore, J.E., Pratt, H.E., Fan, K. et al. An expanded registry of candidate cis-regulatory elements. Nature (2026). https://doi.org/10.1038/s41586-025-09909-9
3. Abdennur Lab., epifolio-notebooks [Source code],[Internet, Accessed: 07 April 2026]. GitHub. https://github.com/abdenlab/epifolio-notebooks
4. Abdennur Lab, epifolio https://abdenlab.org/epifolio/, [Internet, Accessed: 07 April 2026]
5. Agrawal, A., & Scolnick, M. (2025). marimo - an open-source reactive notebook for Python (0.11.26). Zenodo. https://doi.org/10.5281/zenodo.15070030
6. Manz, T., Abdennur, N., & Gehlenborg, N. (2024). anywidget: reusable widgets for interactive analysis and visualization in computational notebooks. Journal of Open Source Software, 9(102), 6939. https://doi.org/10.21105/joss.06939.
7. Manz, T., Gehlenborg, N. & Abdennur, N. (2024). Any notebook served: authoring and sharing reusable interactive widgets. Proceedings of the 23rd Python in Science Conference, https://doi.org/10.25080/NRPV2311

Motivation: Metagenome sequence assembly is an increasingly critical step in studying microbial community composition, and the assembly graphs that this process produces provide a unique window into the underlying genomic diversity thereof. Assembly graphs represent this information at both coarse-grained ("which connected components seem like putative plasmid sequences?") and fine-grained ("what caused this bubble?") levels of detail. Visualizing assembly graphs at both scales -- enabling the user to move back and forth between comprehensive high-level overviews of the graph and clear representations of local details of interest -- remains an active area of research.

Methods: We present MetagenomeScope, a software tool that facilitates the multilevel exploration of assembly graphs. MetagenomeScope iteratively decomposes the graph into interpretable structural patterns and visualizes them as hierarchical annotations. These patterns are used to inform graph layout, providing a clear view of these small-scale details. MetagenomeScope augments this functionality with a rich set of tools for analyzing higher-level graph structure, including path highlighting and interactive summary plots.

Results: MetagenomeScope enables close-up visualizations of densely tangled regions of assembly graphs, such as those caused by rDNA arrays. The tool is able to visualize entire large assembly graphs (up to tens of thousands of nodes) as well as subregions and summaries of even larger graphs.

Significance: The ongoing deluge of sequencing data underscores the need for careful validation. The novel algorithms and features supported by MetagenomeScope simplify assembly graph analysis, lowering the barriers to manual inspection of assembly outputs.

Code and documentation are available at https://github.com/marbl/MetagenomeScope.