Main Figure Reproduction — ALADYNOULLI

Reproducing the Main Manuscript Figures

Each main-text figure in the published manuscript has a dedicated Jupyter notebook in pyScripts/dec_6_revision/new_notebooks/main_paper_figures_clean/ (stripped of large output cells, kept lean for git tracking; full exploration-version notebooks are retained in the parent main_paper_figures/ directory for the authors’ reference). The notebooks run top-to-bottom against the released Zenodo data archive (DOI 10.5281/zenodo.20187989).

Figure	Topic	Notebook	Rendered output
Figure 1	Model overview (architecture schematic + theta/phi/lambda relationships)	`Figure1_Model_Overview.ipynb`	`paper_figs/Fig1.pdf`
Figure 2	Population-level disease signature patterns and temporal evolution	`Figure2_Population_Level_Patterns.ipynb`	`paper_figs/fig2/fig2.pdf`
Figure 3	Individual-specific trajectories (Panels A, B) and heterogeneity within disease subtypes (Panel C)	Panels A + B: `Figure3_Individual_Trajectories.ipynb` Panel C (line-filled heterogeneity plots for MI / Breast / MDD): `R3_Q8_Heterogeneity_MainPaper_Method.html` (already rendered on this site)	`paper_figs/fig3/fig3.pdf` (assembled), `paper_figs/fig3/line_filled_*.pdf` (Panel C panels)
Figure 4	Genetic architecture: GWAS lead variants, RVAS hits, PRS heatmaps per disease subtype	`Figure4_Genetic_Validation.ipynb`	`paper_figs/fig4/fig4.pdf`
Figure 5	Multi-disease risk prediction performance vs PCE / PREVENT / GAIL / QRISK3 / Cox (centered model — as in shipped manuscript)	`Figure5_Predictive_Performance.ipynb` (reparam variant lives under Reparam Results on this site: `Figure5_REPARAM.html`)	`paper_figs/fig5/fig5.pdf`

Shared utility libraries

The notebooks import helper functions from:

pyScripts/fig3_utils.py — Figure 3 trajectory / heterogeneity helpers
pyScripts/fig4_utils.py — Figure 4 AUC / ROC / calibration helpers
pyScripts/fig5utils.py — Figure 5 prediction-evaluation helpers (evaluate_static_auc, evaluate_dynamic_auc, etc.)

Required inputs

Each notebook loads trained checkpoints and population data from the Zenodo deposit:

Model checkpoints (UKB pooled + per-batch): aladynoulli2-main_zenodo.zip → checkpoints/
Numerical data files (signature loadings, lead-SNP tables, performance metrics, RVAS results): SuppDataFiles_for_zenodo.zip
PheCode-ICD mapping (public): phewascatalog.org/phecodes_icd10cm

Individual-level UK Biobank, MGB, and All of Us data are not redistributable; the notebooks can be re-run by approved researchers using their own data access following the procedures described in the manuscript Methods.

Suggested workflow to regenerate any figure

# 1. Clone repo and pull released code archive
git clone https://github.com/surbut/aladynoulli2
cd aladynoulli2
# OR: download the Zenodo v1.0 snapshot
# wget https://zenodo.org/records/20187989/files/aladynoulli2-main_zenodo.zip

# 2. Set up environment (Python 3.12+ with PyTorch, NumPy, Pandas, Matplotlib, Seaborn, scikit-learn)
python3 -m venv aladyn_env
source aladyn_env/bin/activate
pip install torch numpy pandas matplotlib seaborn scikit-learn jupyter

# 3. Open the relevant notebook and run top-to-bottom
jupyter notebook pyScripts/dec_6_revision/new_notebooks/main_paper_figures_clean/Figure2_Population_Level_Patterns.ipynb

Caveats / notes

The figures were rendered against the centered (production) model (pyScripts_forPublish/clust_huge_amp_vectorized.py); the reparam variant (pyScripts_forPublish/clust_huge_amp_vectorized_reparam.py) was the source for Figure 5 panels showing the centered-vs-reparam comparison.
Figure 1 (model overview schematic) is partially hand-assembled in Adobe Illustrator; the notebook produces the component plots used in panels B–D.
Extended Data figures use the same notebook → caption convention. See the Methods section of the manuscript for the full ED-figure pipeline.