BIWT: a bioinformatics walkthrough for embedding spatial multiomics in agent-based models for virtual cells

TL;DR

The BIWT software uses single-cell and spatial data to create virtual cell models that simulate tissue dynamics, bridging static data with dynamic predictions.

Contribution

BIWT introduces a new software tool for initializing spatial agent-based models directly from multiomic data, enabling data-driven virtual cell simulations.

Findings

Initialization strategies in BIWT significantly influence tumor-immune dynamics and spatial clustering in virtual cells.

BIWT provides a framework to generate virtual cells that reflect both experimental and clinical contexts using spatial multiomics data.

Abstract

Whereas transcriptomic and spatial profiling offer static snapshots of tissue structure, mechanistic models use biological rules to predict how tissues evolve. We present the BioInformatics WalkThrough (BIWT) software to directly initialize spatial agent-based models from single-cell and spatial molecular data. We demonstrate how initialization strategies affect tumor–immune dynamics and spatial clustering, positioning BIWT as a software suite to generate data-driven virtual cells representing both experimental and clinical contexts.

The BIWT software is available at https://github.com/PhysiCell-Tools/PhysiCell-Studio. The sample dataset for running the BIWT is available at https://zenodo.org/records/16365625. The code and instructions for reproducing the use case example is available at https://github.com/drbergman/BIWT-Paper.

Figures

Schematic of the BioInformatics WalkThrough and example use case. (A) Users (1) load their data through a graphical user interface, (2) select the cell types to continue with downstream, (3) import the cells into the model, and (4) run the model. (B–M) The BioInformatics WalkThrough spatial-informed initialization produces distinct dynamics in an agent-based model of tumor-immune interactions. (B–D) Well-mixed (B), structured (C), and spatial-informed (D) initialized models. Legend on the right shows the cell type names used in the PhysiCell model. (E–G) Simulation snapshot at endpoint (t = 5 days) for each initialization. (H) Cancer cell population by initialization type over the simulated time. Solid line represents the mean. Shaded area represents ±1 SD. Legend below panel I. (I) Number of ongoing immune attacks at each measured time point. Inset shows these values over the first 6 h. (J–L) Cross-pair correlation function time series heatmaps showing the enrichment or depletion of CD8+ T cells at varying distances from cancer cells. (M) CDF of cancer cell counts in connected components of varying sizes at the final endpoint. Some icons in (A) were generated with Biorender.com.

Full text

1 Introduction

Advances in sequencing technologies have revolutionized our understanding of cellular biology, offering unprecedented insights into cell behavior and interactions (Ståhl et al. 2016, Marx 2021). Coupled with bioinformatics tools, we can now profile cells in their spatial context (Williams et al. 2022), classify distinct cell types (Alquicira-Hernandez et al. 2019, Zhang et al. 2023), and infer active signaling pathways (Cabello-Aguilar et al. 2020, Cherry et al. 2021, Jin et al. 2025)—providing a more comprehensive view of tissue organization and function. These advances not only illuminate biological structure but also motivate the development of virtual cell models that predict how tissues evolve over time.

Agent-based modeling (ABM) addresses this need by simulating the behavior of individual cells—represented as agents that interact with each other and their environment according to mathematically defined rules (An et al. 2009). This rule-based approach enables in silico exploration of complex, multicellular systems over time, but building and calibrating such models typically depends on expert knowledge and experimental data. Incorporating transcriptomic insights can introduce cell-type-specific parameters and pathway activity, enhancing model accuracy, utility, and reproducibility (Arulraj et al. 2024). However, even the first step—initializing agent-based models, particularly in space—remains a critical challenge, as different starting conditions can yield dramatically different outcomes. Despite advances in both bioinformatics and modeling frameworks, bridging these domains in a way that is accessible, reproducible, and spatially aware remains an open challenge in computational biology.

Because single-cell and spatial molecular technologies measure the phenotype of individual cells, the data generated from their analyses perfectly mirror the variables tracked in ABM. This close relationship between computational models and high-throughput data makes ABMs primed for initialization from single-cell and spatial molecular data. Here, we present the BioInformatics WalkThrough (BIWT), an open-source tool that integrates bioinformatics analyses into the ABM framework PhysiCell (Ghaffarizadeh et al. 2018).

2 Results

2.1 Implementation

BIWT provides a Python-based graphical user interface (GUI) that guides users through a streamlined process for converting transcriptomic data into initial conditions for PhysiCell simulations. The workflow proceeds stepwise—from dataset selection and cell type prioritization to data-driven generation of model inputs (Fig. 1; see Methods and Figs S1–S16, available as supplementary data at Bioinformatics Advances online for details). BIWT supports commonly used formats from both R and Python pipelines, including Seurat (Hao et al. 2021), SingleCellExperiment (Amezquita et al. 2020), and AnnData (Wolf et al. 2018), as well as CSVs. For non-spatial data, BIWT assigns cells based on their relative abundances in spatial distributions specified by the user. For spatial transcriptomics data, it generates a spatially resolved virtual microenvironment that perfectly matches the spatial distribution of real cells in the measured tissue.

Schematic of the BioInformatics WalkThrough and example use case. (A) Users (1) load their data through a graphical user interface, (2) select the cell types to continue with downstream, (3) import the cells into the model, and (4) run the model. (B–M) The BioInformatics WalkThrough spatial-informed initialization produces distinct dynamics in an agent-based model of tumor-immune interactions. (B–D) Well-mixed (B), structured (C), and spatial-informed (D) initialized models. Legend on the right shows the cell type names used in the PhysiCell model. (E–G) Simulation snapshot at endpoint (t = 5 days) for each initialization. (H) Cancer cell population by initialization type over the simulated time. Solid line represents the mean. Shaded area represents ±1 SD. Legend below panel I. (I) Number of ongoing immune attacks at each measured time point. Inset shows these values over the first 6 h. (J–L) Cross-pair correlation function time series heatmaps showing the enrichment or depletion of CD8+ T cells at varying distances from cancer cells. (M) CDF of cancer cell counts in connected components of varying sizes at the final endpoint. Some icons in (A) were generated with Biorender.com.

2.2 Spatial initialization drives emergent behavior in ABM of tumor-immune dynamics

To illustrate BIWT’s utility, we applied it to an agent-based model of tumor-immune interactions composed of malignant epithelial cells, healthy epithelial cells, and CD8+ T cells (see Methods for details). The model tests how the spatial location of immune infiltration influences tumor progression. Therefore, in the model only cancer cells proliferate or undergo apoptosis; CD8+ T cells migrate and kill cancer cells while healthy epithelial cells neither divide nor die.

We initialized the model using spatial transcriptomics data from a recent clinical trial of neoadjuvant immunotherapies for pancreatic ductal adenocarcinoma (PDAC) (Li et al. 2022, Sidiropoulos et al. 2024). In all cases, the abundance of each cell type was taken from a surgical biospecimen treated with the triple combination of GVAX, anti-PD1, and CD137 agonist therapies. This sample was selected from the cohort of trial samples as it exhibited high immune cell density at the border of the malignant compartment, making it a compelling test case for BIWT. We compared three initialization strategies of the ABM: a well-mixed initialization, a structured initialization, and a spatial-informed initialization (Fig. 1B–D). In the well-mixed initialization, all three cell types are uniformly distributed in a disc. In the structured initializations, immune cells surround epithelial cells in an annulus. The spatial-informed initialization maps each cell directly from the spatial transcriptomics data, preserving the original coordinates and local microenvironmental structure. Each initialization was simulated three times over five simulated days, yielding nine runs in total (Fig. 1E–G). Supplementary movies (S1–S3, available as supplementary data at Bioinformatics online) visualize the progression of each scenario over time.

To assess the impact of BIWT’s spatial-informed initialization on tumor dynamics, we tracked cancer cell counts over time (Fig. 1H). Each initialization produced distinct trajectories. The model with the well-mixed cells had a sharp early decline in cancer cells before stabilizing at 90% of the initial population. The structured and spatial-informed setups followed similar trends but diverged in tumor burden, with the spatial-informed case yielding significantly higher cancer cell counts. Notably, the spatial-informed setup exhibited greater variability in cancer cell counts at early time points across replicates—even though cell placement was identical across runs—whereas the structured initialization, which varied placement, produced more consistent outcomes (Fig. S17, available as supplementary data at Bioinformatics Advances online). This counterintuitive result underscores a core property of spatial ABMs: their intrinsic stochasticity enables a single initialization to produce a distribution of outcomes, and this distribution is sensitive to the initial conditions.

We next examined immune activity across the three initialization strategies by tracking ongoing immune cell attacks over time. The well-mixed condition exhibited a sharp, early spike in immune engagement (Fig. 1I), consistent with the initial drop in cancer cell counts observed in this setup (Fig. 1H). Given this relationship, we anticipated lower immune activity in the structured and spatial-informed models, which showed early tumor growth. Nonetheless, the spatial-informed setup exhibited more immune engagements than the structured case, despite faster cancer expansion. Over the full simulation, it also accumulated more immune cell engagements than the structured initialization (Fig. S18, available as supplementary data at Bioinformatics Advances online).

To assess how initialization influences immune–tumor colocalization over time, we computed the cross-pair correlation function (cross-PCF) between cancer cells and CD8+ T cells. The cross-PCF quantifies whether T cells are enriched (>1) or depleted (<1) at varying distances from individual cancer cells. We averaged these values across cancer cells at each time point and visualized them as vertical strips in Fig. 1J–L (see Methods, available as supplementary data at Bioinformatics online). In the model with well-mixed initialization, immune cells become progressively depleted near cancer cells, likely reflecting successful killing of initially colocalized targets (Fig. 1J). In contrast, the models with structured and spatially informed initializations show progressive immune infiltration, evident in the downward shift of T cell enrichment toward shorter cancer—T cell distances over time (Fig. 1K–L). These divergent patterns illustrate how initialization governs spatial interactions between cancer and immune cells.

Finally, we quantified cancer cell clustering at the final time point to assess how initial conditions shaped tumor architecture. Endpoint snapshots (Fig. 1E–G) reveal visually distinct cancer cell groupings. We quantified these using connected components derived from neighbor graphs in the PhysiCell output (see Methods; Fig. S19, available as supplementary data at Bioinformatics Advances online). We then computed the cumulative distribution function (CDF) of cancer cell counts per component, averaging across replicates and shading 1 SD from the mean (Fig. 1M). In the well-mixed initialization, nearly all cancer cells end in small clusters ( $[eqn]$ 100 cells). In contrast, more than half of the cancer cells in the structured model form large clusters ( $[eqn]$ 1000 cells). The spatially informed models show an intermediate pattern: most cells belong to large clusters, but around 40% remain in smaller ones. These differences reinforce the importance of initialization in shaping emergent tissue architecture and demonstrate the benefit of using real spatial data to initialize ABMs.

3 Conclusion

Integrating biological data directly into mathematical models is a longstanding goal in agent-based modeling (Lorenzo et al. 2024). BIWT advances this aim within the PhysiCell framework by enabling users to build models through a graphical interface. By streamlining the creation of data-driven initial conditions, BIWT allows researchers to explore patient-specific dynamics in a reproducible and accessible way (Mangul et al. 2019). Our results show that spatial dynamics—such as immune infiltration, tumor compactness, and cell clustering—can significantly influence simulated outcomes. These findings make clear that initialization is not a minor technical detail, but a central design variable that shapes model behavior and interpretability.

Specifically, our simulations represent the competition of tumor cells for limited resources, modeled here as physical space. As a result, the initial spacing of cancer cells has a strong influence on subsequent growth. Over five simulated days, these initialization effects remain measurable and could lead to conflicting interpretations in downstream analyses if not explicitly addressed. BIWT mitigates this risk by grounding initialization in real spatial data, reducing ambiguity and modeler bias in simulations aimed at digital twin applications. We expect that the increased variability observed in simulations initialized from real spatial distributions reflects biological uncertainty in patient outcomes. This underscores the need for future work in uncertainty quantification and in identifying which additional measurements—such as spatial patterns at later time points—would most improve model predictivity. Extending such studies from individual patients to full cohorts will also require reproducible pipelines that support scalable data integration and simulation setup across multiple samples.

BIWT’s modular architecture supports integration with a wide range of data types and modeling frameworks. On the input side, it accommodates standard formats from both R and Python single-cell analysis workflows, ensuring compatibility with evolving data pipelines. On the output side, while BIWT currently targets the PhysiCell framework, its design enables adaptation to other agent-based modeling platforms such as Chaste (Mirams et al. 2013) and CompuCell3D (Swat et al. 2012), which support distinct mathematical formalisms suited to different biological questions. This interoperability positions BIWT as a flexible bridge between molecular measurements and mechanistic modeling. While BIWT enables model initialization by cell type, future work will extend this platform to automate the assignment of phenotype-specific parameters such as proliferation rate, receptor expression, or metabolic state from atlas studies to enable context-specific model design.

The BioInformatics WalkThrough (BIWT) is the first tool of its kind to bridge bioinformatics pipelines and agent-based modeling through an intuitive, GUI-driven interface. By anchoring spatial agent-based simulations in real molecular data, BIWT supports reproducibility, accessibility, and biological realism. As agent-based models evolve into predictive digital twins for translational research, BIWT provides the infrastructure to transform static molecular profiles into dynamic, interpretable, and patient-specific simulations.

Supplementary Material

btaf571_Supplementary_Data