QUICHE simulation design

Jolene Ranek; Mike Angelo

Nov 26, 2025

QUICHE simulation design

DOI

https://dx.doi.org/10.17504/protocols.io.6qpvrwjk2lmk/v1

Jolene Ranek¹,
Mike Angelo¹

¹Department of Pathology, Stanford University, Stanford CA, USA

Jolene Ranek

DOI: https://dx.doi.org/10.17504/protocols.io.6qpvrwjk2lmk/v1

Protocol Citation: Jolene Ranek, Mike Angelo 2025. QUICHE simulation design. protocols.io https://dx.doi.org/10.17504/protocols.io.6qpvrwjk2lmk/v1

Manuscript citation:

Ranek, J. S., Greenwald, N. F., Goldston, M., Fullaway, C. C., Sowers, C., Kong, A., Mouron, S., Quintela-Fandino, M., West, R. B., and Angelo, M. QUICHE reveals structural definitions of anti-tumor responses in triple negative breast cancer. bioRxiv. 2025

License: This is an open access protocol distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited

Protocol status: Working

We use this protocol and it's working

Created: November 26, 2025

Last Modified: November 26, 2025

Protocol Integer ID: 233510

Keywords: quiche simulation design, quiche simulation design this protocol, triple negative breast cancer, simulation design for the manuscript, simulation design, automated computational, quiche

Abstract

This protocol describes the simulation design for the manuscript "The automated computational workflow QUICHE reveals structural definitions of anti-tumor responses in triple negative breast cancer" by Ranek et al.

In silico simulation design

To validate our approach and benchmark spatial enrichment methods on their ability to identify condition-specific differences in the spatial organization of cell types from individual patient samples, we simulated multi-patient cohorts with different spatial topologies (e.g., unstructured, structured), where the ground truth organization was known. Here, each multi-patient cohort consisted of 20 patient samples evenly split across two conditions. For each patient sample, we fixed the number of cells and cell types to be the same, and then varied the spatial organization of cells according to each patient's condition label (e.g., condition A consisted of immune cells enriched in the cancer core). 

Unstructured simulation

We generated simulated datasets with unstructured spatial topologies as follows. For each patient sample, we randomly sampled the spatial coordinates of 5000 cells across five cell types (cell types A, B, C, D, E) from a uniform distribution within a     pixel domain. Each image was then discretized into a   grid. To define ground truth condition-specific niches, we selected one grid from each image and enforced cell types to co-occur within that region (condition 1: cell types A, C, and E; condition 2: cell types B and D). The grid size   was used to determine the niche prevalence (e.g.,   grid indicates that a niche occupies   or   of a sample). The output of this approach is a ground truth cell-to-cell type and cell-to-niche assignments. To evaluate how well methods can identify condition-specific differences, we varied both the size of a niche in a patient sample (grid size = 3, 4 , 5, 6, 7, 8, 9, 10) and its prevalence across patient samples (20, 40, 60, 80, 100% of samples). Method performance was assessed over five random trials for each parameter configuration.

Structured simulation

For a more realistic simulation design, we leveraged multiplexed proteomic imaging data from the Spain TNBC cohort to define cell spatial coordinates and tissue domains as follows.

Tumor selection: To generate simulated datasets with structural spatial topologies, we randomly selected tumors with compartmentalized tumor-immune spatial architectures using the mixing score defined in Ref. [1] as, 

The mixing score quantifies the degree of mixing between reference (tumor) and target (immune) cells within a fixed spatial radius,  , by computing the ratio of heterotypic (cancer-immune) to homotypic interactions (cancer-cancer, immune-immune). Here,   and   represent the number of reference and target cells, respectively. The indicator functions   and   denote whether a target or reference cell type is within the neighborhood  of cell  . Specifically,   if cell   is a target cell type within the neighborhood  , and   otherwise. Similarly,   if cell   is a reference cell type within the neighborhood  , and   otherwise. The neighborhood   is defined as  , where   and   are the spatial coordinates of cells   and  , respectively. Tumors were considered to have compartmentalized spatial architectures if they had more homotypic than heterotypic interactions.

Tumor compartments: To define structural regions in each tumor image, we followed the approach outlined previously [2]. Briefly, tumor regions (cancer core, cancer border, stroma) were defined by taking the union of the E-cadherin signal with the cancer cell segmentation masks. Tumor compartment masks were then thresholded (0.0015), binarized, and eroded by 100 pixels to delineate cancer cells within the core vs border; all of the remaining cells were defined as stroma. Image processing was performed using the scikit-image v0.19.3 package in Python.

Differential spatial enrichment: We simulated condition-specific cell type enrichment within structured spatial topologies (condition 1: immune cells within the cancer core, condition 2: immune cells within the cancer border). For each patient sample, we randomly sampled cells from tumor regions (condition 1: cells within the cancer core; condition 2: cells within the cancer border) using a Gaussian distribution with a specified radius (radius = 50, 100, 250, 500). Next, we defined ground truth niches by reclassifying a subset of cells within the Gaussian radius as immune cells, thereby creating localized clusters of immune infiltration. Finally, we varied the (1) niche size (radius = 100, 250, 500 pixels), (2) niche sparsity (5, 10, 25% of immune cells), and (3) niche prevalence across patient samples (20, 40, 60, 80, 100% samples). Method performance was evaluated across five random trials for each parameter configuration.

Expression

We used Splatter [3] to simulate expression data in our in silico data. Simulation parameters were estimated from a single-cell RNA sequencing dataset [4]. Next, the estimated parameters (mean_rate = 0.0173, mean_shape = 0.54, lib_loc = 12.6, lib_scale = 0.423, out_prob = 0.000342, out_fac_loc = 0.1, out_fac_scale = 0.4, bcv = 0.1, bcv_df = 90.2, dropout = None) were used in the Splatter groups function (python wrapper scprep SplatSimulate v1.2.3 of splatter v1.18.2) to simulate expression data (p = 500) for each cell type (e.g., cancer core, cancer border, stroma, immune).

References

[1] Keren, L., Bosse, M., Marquez, D., et al. A structured tumor-immune microenvironment in triple-negative breast cancer revealed by multiplexed ion beam imaging. Cell 174(6):1373–1387.e19, 2018. 

[2] Greenwald, N.F., Nederlof, I., Sowers, C., et al. Temporal and spatial composition of the tumor microenvironment predicts response to immune checkpoint inhibition. bioRxiv 2025.01.26.634557, 2025.

[3] Zappia, L., Phipson, B., & Oshlack, A. Splatter: simulation of single-cell RNA sequencing data. Genome Biol. 18(1):174, 2017.

[4] Loh, K.M., Chen, A., Koh, P.W., et al. Mapping the pairwise choices leading from pluripotency to human bone, heart, and other mesoderm cell types. Cell 166(2):451–467, 2016.