pdac_tumoroid

Code accompanying:

An integrated platform for high-throughput phenospace learning of 3D multilineage organoid systems Okuda R, Harmel C, Xu Q, Mary H, Schulz P, Steinacher L, D'Arcangelo E, Gjeta B, Signer M, Cubela I, Bickle M, Lutolf MP, Cabon L, Lukonin I, Camp G. bioRxiv (2025). doi:10.1101/2025.11.27.690952

This repository contains the experimental–computational pipeline for high-throughput spatial phenotyping of 3D multilineage organoids. The system pairs a 384-well pancreatic ductal adenocarcinoma (PDAC) + cancer-associated fibroblast (CAF) tumoroid culture with multiplexed imaging, then learns a quantitative "phenospace" that preserves spatial architecture. We use it to identify pathway modulators of the fibrotic microenvironment and stroma-dependent vulnerabilities, and extend the platform to immune-competent systems and patient-derived explants.

Overview

The pipeline runs in three stages:

1. Image processing (image_processing/) — turn raw multiplexed microscopy into single-object features.
2. Phenospace learning (phenocoder/, laminator/) — embed objects into a learned latent space and characterize spatial organization.
3. Analysis & figures (analysis/) — reproduce the screens, benchmarks, and figures from the paper.

Configuration is YAML-driven (configs/), and the heavy compute steps are dispatched to HPC schedulers via job scripts (bsub/ for LSF, sbatch/ for Slurm).

Setup

The full software environment (Python + R/Bioconductor) is frozen in environment.yml:

conda env create -f environment.yml   # creates the `pdac_tumoroid` environment
conda activate pdac_tumoroid

Run scripts from the repository root so the local packages (analysis, phenocoder, image_processing, benchmarking, laminator) are importable. GPU (CUDA) is required for Cellpose segmentation and CVAE training.

Repository layout

Path	Description
image_processing/	CLI-driven microscopy pipeline: flatfield & illumination correction, background correction, MIP, 3D stitching, cross-cycle registration, Cellpose segmentation, feature extraction, neighborhood construction, montaging, and QC.
phenocoder/	Core ML package (also see standalone implementation of Phenocoder). Conditional/standard convolutional VAE, patch generators for nuclei or grid sampling, phenospace embedding, spatial graphs & convex-hull metrics via squidpy, clustering, training, and plotting.
laminator/	Reimplementation of Laminator (Wahle et al., Nat. Biotechnol. 2023) for "laminating" organoid images into oriented windows and radial-neighborhood message passing. Used for length-scale analyses.
benchmarking/	Embedding-quality metrics: graph connectivity, multi-label AMI, cLISI/CAS/NASW, plus ARI/NMI, used to compare phenocoder embeddings against nuclei-only and alternative baselines.
configs/	Per-screen YAML configs and master parameters (screens, channel/marker maps, plates, phenocoder model params).
bsub/, sbatch/	Cluster job scripts for image processing, dataset generation, model training, phenocoding, and benchmarking.
analysis/	Numbered, per-experiment pipelines that reproduce the study (see below).

Analyses

Each subdirectory of analysis/ is a self-contained, numerically ordered pipeline (01_…, 02_…) mixing Python and R. Files prefixed run_ perform computation; files prefixed plot_ generate figures. Subdirectories are listed in the order their main figures appear in the paper.

Subdirectory	Corresponds to
brightfield_imaging/	Brightfield imaging montages.
z_undersample_and_segmentation_eval/	Validation: z-stack undersampling and segmentation-quality assessment.
length_scale_sensitivity/	Laminator-based message passing and distance-scale sensitivity analysis.
timecourse/	Multilineage developmental timecourse — diffusion maps, MLP marker imputation, spatial prediction, latent-space traversal.
pilotscreen/	FDA-compound pilot screen, including cross-plate label transfer and spatial-graph rendering.
multiple_patients/	Generalization across multiple patient-derived lines.
tumoroidscreen/	Main PDAC + CAF inhibitor screen — phenocoding, benchmarking, feature heatmaps, compound-effect analysis, STRING network, example tumoroids.

Data availability

The processed images and mdata (.h5mu) files are deposited in the EMBL-EBI BioImage Archive:

DOI: TBD — add accession/DOI on publication.

Each screen is organized as <screen>/images/ and <screen>/mdata/. The analysis scripts read these .h5mu files (see analysis/).

License

MIT — see LICENSE.md. © 2026 DevSystems Lab.