Skip to content

Plotting age of mini-brain #3

Open

Description

@Liberty23513

Title

Organoid growth charts: normative modelling of brain-organoid cell-type composition

Leaders

Boting Li
liboting940523@gmail.com
Discord: boting_90337

Collaborators

Open. One or two people from the Oct 2 normative modelling tutorial could be a better fit.

Project Description

Brain organoids are mm-sized pieces of human brain tissue grown from stem cells, which develop many of the cell types of the fetal cortex in a dish. The longer they are kept in culture, the older they become biologically: organoids of a few weeks resemble first-trimester fetal cortex, and organoids kept for over a year start to show postnatal-like features. Because they are human and can be grown from patients, they are increasingly used to study brain development and disease.

Paediatricians read a child's height off a growth chart: age on the x-axis, height on the y-axis, centile bands drawn from thousands of healthy children. Brain organoid labs have no equivalent. People say an organoid batch "looks off" without being able to say how far off, or off compared to what.

This project builds that chart. X-axis is days in culture, y-axis is the fraction of a given cell type (starting with neural progenitors), and the reference set is the public Human Neural Organoid Cell Atlas, which its authors already curated to exclude diseased, mutation-carrying and hormone-treated samples. The approach borrows from normative modelling, the family of methods behind "brain charts" in neuroimaging, with organoid age in place of subject age and lab of origin in place of scanner site.

Two outputs. First, the chart itself. Second, and more useful: by refitting the chart for the largest protocol (131 samples) on random subsets of 20, 40, 80 and all 131, a curve showing how many organoids a lab actually needs before the centiles are stable. That number is currently guessed.

Link to project repository/sources

https://github.com/Liberty23513/organoid-growth-charts

Goals for Brainhack Montreal

Everyone: start from the ready-made per-sample composition table, check the cell-type groupings, and plot the raw scatter per protocol. Beginner-friendly, no modelling needed.

Target: fit a normative model of how the progenitor fraction changes with organoid age (age is the input, progenitor fraction is the measured outcome), with lab of origin as the site effect, and produce the centile chart. Fallback if the standard toolkit does not take proportional data: quantile regression, which gives a comparable figure in far less code.

Stretch: the subsampling curve (how many samples are needed for stable centiles), extending to several cell types on a CLR scale, and extending the Velasco-protocol chart from 6 months to 5 years with the 34 long-term organoids from Faravelli et al. 2026 (GEO GSE333707). Scoring a disease-organoid dataset against the healthy reference is a further option if time allows.

Skills

  • Python with pandas and a plotting library is enough to contribute to the first goal.
  • Normative modelling or any experience with PCNtoolkit, GAMLSS, quantile regression or centile estimation. Having taken the Oct 2 tutorial counts.
  • Single-cell data handling (anndata, scanpy) only if you want to rebuild the table at a different cell-type resolution.
  • Non-coding and genuinely needed: judging which cell-type labels are comparable across studies, and which differentiation protocols should be pooled versus kept separate. This sets the ceiling on everything downstream.

Onboarding documentation / Expectations

Background reading, none of it required before arriving:

  • HNOCA atlas paper: He, Dony, Fleck et al., Nature 635, 2024, doi 10.1038/s41586-024-08172-8.
  • PCNtoolkit (Predictive Clinical Neuroscience Toolkit), v1.3.0 on PyPI: https://pypi.org/project/pcntoolkit/
  • Alyas, Horsley, Little, Taylor, Wang, Leiberg, "Normative Modelling in Neuroimaging: A Practical Guide for Researchers", arXiv:2509.07237 (2025).
  • Closest organoid work: Faravelli, Antón-Bolaños, Wei et al., "Human brain organoids record the passage of time over multiple years", Nature 657, 713-723 (2026), doi 10.1038/s41586-026-10877-x. Same cortical protocol as our largest group, followed to 5 years; it reports composition over time and replicate variability, but not reference centiles.

The per-sample composition table (322 samples, one row each: age, protocol, publication, cell line, cell-type fractions) is already built and sits in the repo as data/hnoca_sample_composition.csv, together with the script that rebuilds it from the public atlas. This CSV is the working data for the whole hackathon, so day one starts on a 170 KB file rather than on the 19 GB atlas. Everything after that runs on a laptop in seconds. Come with questions about whether this framing holds up; the modelling choices are genuinely open.

What will participants learn?

  • How to apply normative modelling outside neuroimaging, including what breaks when the response variable is a proportion rather than a volume.
  • How to decide how much reference data a centile model needs, by measurement rather than by rule of thumb.
  • Working with public single-cell atlas metadata, and how much harmonization it still needs.

Data to use

Human Neural Organoid Cell Atlas (HNOCA), He, Dony, Fleck et al., Nature 2024, doi 10.1038/s41586-024-08172-8. Public on CELLxGENE. It pools data from 36 datasets produced by many labs, re-annotated with one common set of cell-type labels: 1.77 million cells, 26 differentiation protocols (3 unguided, 23 guided), ages 7 to 450 days, 396 batches at the dataset-plus-sample level. Cell-type annotations are included, and the authors removed diseased, mutation-carrying and hormone-treated samples, which is exactly what a normative reference needs.

Protocols are named after the paper that published them. Two carry most of the data and are used by several labs: the Velasco 2019 dorsal cortex protocol (131 samples, 4 publications, 11 cell lines, days 21 to 192) and Lancaster 2014 (81 samples, 3 publications, 8 cell lines, days 9 to 120). We build the charts within those two, so lab of origin can be modelled as a site effect separately from protocol. The other 25 protocols mostly come from a single publication each (22 of 27 overall), so for them protocol and lab cannot be separated; we report that rather than pretend otherwise.

Credit to collaborators

Contributors listed in the README via the all-contributors bot, including non-coding contributions such as cell-type harmonization and figure review. If this becomes a preprint, contributors are co-authors, and we agree authorship order at the wrap-up session rather than months later.

Image

Progenitor fraction against organoid age for the Velasco and Lancaster protocols, coloured by lab.

Image

Type

methods development, pipeline development, visualization

Project Maturity Status

1 - prototype

Topic

bayesian_approaches, data_visualisation, hypothesis_testing, reproducible_scientific_methods, statistical_modelling

Tools

Jupyter, Machine_learning, other

Programming language

Python

Modalities

other

Git skills

1 - commit and push, 2 - branches and pull requests

Anything else?

Something else needed:

A laptop with Python 3. What to install depends on which part you want to work on:

Explore and plot the data table: pip install pandas matplotlib. Basic pandas is the only skill needed.
Fit the growth chart: also pip install pcntoolkit statsmodels. Some regression or normative modelling experience helps.
Rebuild the data table from the full atlas (rarely needed): also pip install remfile h5py anndata. Familiarity with single-cell data helps.
Cell-type and protocol judgement calls: no install, just organoid or developmental neuroscience knowledge.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions