1 of 10

Brain Organoid Morphological Classifier

Automated QC for stem-cell-derived brain organoids

Shiraz Bheda | BioAgent BOMC | June 2026

2 of 10

Context

Human-relevant models need trustworthy QC

3 of 10

Why neuroscience needs better models

  • Neurodegenerative and neurodevelopmental disorders involve human-specific brain biology that is difficult to model.
  • Therapeutic target discovery depends on systems that recapitulate human development and pathology — not just rodent proxies.
  • Mouse models drove decades of discovery but often miss human cortical architecture, timing, and genetic context.
  • NIH is steering the field toward human-relevant platforms — including iPSC-derived organoids — to improve translational confidence.

​

4 of 10

Brain organoids — promise & pitfall

  • 3D stem-cell-derived brain organoids capture aspects of corticogenesis and model microcephaly and related disorders.
  • But protocols yield variable morphology; subjective manual grading injects noise into omics, screening, and target ID.

5 of 10

Our solution

Literature-informed morphological QC

6 of 10

Organoid-Grader

  • Automated pass/fail classifier for brightfield brain organoid images.
  • Validated on Schroter et al. (Zenodo 10301912): ~1,400 images, 64 organoids, 2 labs.
  • Gates organoids before costly downstream assays.

7 of 10

Image classifier + literature curation

  • Extracts quantitative morphology from images (with ground-truth masks for training).
  • Literature review layer prioritizes features with strongest QC evidence — necrosis, cysts, size, uniformity.
  • Random Forest learns from rule-derived labels; not a black-box CNN alone.

8 of 10

Priority features (must all pass)

  • Brightness / texture uniformity
  • Necrotic core (strict)
  • Max Feret diameter — neural tissue proxy
  • Cyst formation ratio

9 of 10

Pipeline & results

  • Feature extraction → literature-weighted grader → ML classifier.
  • GroupKFold CV by organoid ID; tuned thresholds yield balanced pass/fail on 1,407 images.
  • CLI, Streamlit UI, and demo video for submission.
  • Two major iterations
    • 1st iteration: agnostic feature selection (quantifying organoid criteria relative to median distribution; 100% passed)
    • 2nd iteration: literature-informed feature selection and prioritization (32% passed)

10 of 10

Next iteration: epigenetic QC

  • Integrate DNA methylation profiling to build methylation clocks on donor cell lines.
  • Verify biological age signatures are truly reset before differentiation — a prerequisite for trustworthy organoid biology.
  • Accurate clock analysis builds confidence when modeling neurodegenerative and neurodevelopmental disorders, where epigenetic state can confound phenotypes.
  • Vision: morphological QC + methylation clock QC as a combined gate before mechanistic studies.