tutorial_HSC

MorphAgent HSC tutorial — Figure 3

Recorded walkthrough: tutorial_HSC.mp4 (2 min) runs the notebook end to end.

One notebook. It replays the Young/Old PCA panels from:

• HSC_Fig3a_YoungOld_reproduction_20260916.zip (110-cell discovery)
• HSC_Fig3a_YoungOld_reproduction_20260916 2 (162-cell confocal transfer and blur)

Copied under source/fig3a_bundle/.

tutorial_HSC/
  notebook/hsc.ipynb
  source/fig3a_bundle/
    input/selected_features.csv              locked 25 names
    input/features_code_filtered_*.csv       110-cell MorphAgent features
    input/features_manuel_*.csv              110-cell handcrafted features
    input/omics_aligned_*.tsv                paired transcriptome
    input/merged_numeric_features_cleaned.csv  162-cell confocal transfer
    input/merged_blurred_features.csv          same 162 cells, blurred
    scripts/
  data/outputs/

Setup

python -m pip install -r requirements.txt

Open notebook/hsc.ipynb and run all cells. Same functions as source/fig3a_bundle/run_all.sh (blur panel uses --combo-index 11, the published 0.795 setting).

Expected CV ROC AUC (5-fold × 5 repeats, seed 42)

Analysis	n	Paper	This notebook
MorphAgent, fixed 25 image features	110	0.739	0.739
Handcrafted, two features, polynomial degree 3	110	0.575	0.575
Transcriptome, top 500 of ranked 2,000	110	0.920	0.920
Transfer, same 25 features, confocal	162	0.905	0.905
Blurred 162 cells, combo 011	162	0.795	0.795

The scatter plot is fitted on all cells in that panel; its in-sample full_roc_auc is not the reported CV AUC.

The 25-name list is taken as a given. All 25 names are columns of the discovery and transfer tables. This notebook evaluates that list; it does not re-select features.

Combo 011 is standard scaler, clip quantile 0.02, signed log1p, no PCA whitening. Running the full 36-combination blur grid yields a higher best AUC (~0.835); that is not the published panel.

CLI

PYTHON_BIN=python bash source/fig3a_bundle/run_all.sh data/outputs