organoid_inference

Bayesian inference of organoid morphogenesis parameters from single static microscopy images, using a multi-cell phase-field model, the SampEuler shape descriptor (a sampled Euler characteristic transform), and approximate Bayesian computation (ABC) by rejection.

This repository contains the code needed to regenerate the simulations, descriptors, and posteriors. The experimental microscopy images used as inference targets are in data/experimental_images/; all simulator output and intermediate caches are regenerated by the scripts below rather than shipped.

Method overview

We infer the triple (tauV, xi, t): the cell-volume relaxation time, the lumen osmotic pressure, and the developmental time, from a single snapshot of unknown developmental age:

1. Draw (tauV, xi) uniformly from the prior and simulate the phase-field model forward. Each run writes a frame at a fixed model-time cadence until its (parameter-dependent) stopping criterion is met. Pooling every frame across all runs gives a candidate set that is uniform over the reachable region {(tauV, xi, t) : 0 <= t <= T(tauV, xi)}.
2. Featurise every candidate and each observation with the SampEuler descriptor.
3. ABC rejection: for each observation, accept the candidates closest under the SampEuler–Wasserstein distance. The candidate pool is built once and reused across all observations.

Repository layout

scripts/
  abc_rej/                      ABC-rejection pipeline (the main method)
    generate_sims.py              Stage A1: draw the prior grid, simulate, cache metadata
    sample_and_featurise.py       Stage A2: pool-sample frames, compute SampEuler descriptors
    compute_distances.py          Stage B: SampEuler–Wasserstein distances + acceptance
    featurise_temporal.py         Trajectory (multi-frame) featurisation
    featurise_diamnorm.py         Size-normalised featurisation for the experimental targets
    observations.py               Load and featurise the experimental images
    exp_distances.py              Distances and acceptance for the experimental targets
    sim_worker.py, config.py      Per-theta workers and configuration
    plot_posterior.py, plot_posterior_3p.py, plot_exp_corner.py
  core.py                       Simulator wrapper, ECTComputer, Wasserstein distance
  sampeuler.py                  SampEuler descriptor (sampled Euler characteristic transform)
  paths.py                      Project path configuration (edit for your machine)
  launch_abc_rej_gen.sh         Stage A1 dispatch across hosts
  launch_abc_rej_featurise.sh   Stage A2 dispatch across hosts
  launch_abc_rej_dist.sh        Stage B dispatch, one observation per host
  launch_abc_rej_temporal.sh    Trajectory-cache dispatch across hosts
  compute_sampeuler_bounds.py   Regenerate data/sampeuler_bounds.npz (run first)
  sweep.py                      (tauV, xi) sweep driver for the phase-field simulator
  gen_fakexp_one.py             Generate a synthetic validation organoid
  recovery_wasserstein_cuda.py  GPU frame-recovery under the Wasserstein distance
  abc_rej_distance_cuda.py      GPU SampEuler–Wasserstein distance computation
  regress_morphometrics.py      Ridge regression of morphometrics on SampEuler
  plot_recovery_cdf.py          Cumulative recovery-error distribution
  plot_recovery_per_param.py    Per-parameter recovery error versus true time
data/
  experimental_images/          Experimental organoid microscopy (.tif), inference targets

Requirements

• Python 3.9+
• numpy, scipy, pandas, scikit-learn, matplotlib, seaborn, Pillow, tqdm, numba, joblib (see requirements.txt).
• eucalc, the Euler-characteristic library used by the descriptor. It is made importable through scripts/paths.py:ensure_eucalc(), which adds a local lib/ directory to the path; build/install eucalc and point that helper at it.
• MCPFM phase-field solver (C++), not included here. It is the model of Tanida et al. (2025) and Lee et al. (2026); set scripts/paths.py:MODEL_REPO to your local build.
• A CUDA GPU is required only for the *_cuda.py scripts.

Running the code

This repository is the code behind the published results. It was run across a small cluster with shared storage, so the paths and host allocations are specific to that setup. Adapt them before running anything:

• scripts/paths.py for the project, data and library locations
• scripts/abc_rej/config.py for the cache and output directories
• the HOSTS arrays in scripts/launch_abc_rej_*.sh, and ABC_REJ_REMOTE_DIR, for the machines that jobs are dispatched to

License

Copyright (C) 2026 Haochen Yang.

This project is licensed under the GNU General Public License v3.0 or later (GPL-3.0-or-later), the same licence as the sampeuler package. You may redistribute and/or modify it under the terms of the GNU GPL as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version; see LICENSE. GPL is a copyleft licence: any distributed work derived from this code must also be released as open source under the same terms.