Satyawan1/alzheimer-research-complete

Alzheimer's CascadeNet — Multi-Scale Neuroscience Discovery

7 novel from-scratch models trained on 4.7 GB of real neuroscience data from 15+ public sources. All models 100% from scratch. Zero pretrained weights. Novel architectures.

Models

#	Model	Params	AUC	File	Key Innovation
1	CascadeNet	141K	0.727	models/best_cascadenet.pt	Biologically-informed amyloid cascade
2	BloodCascadeNet	298K	0.733	models/best_blood_cascadenet.pt	Blood-only AD screening (no scans)
3	Temporal CascadeNet	680K	0.958	models/best_temporal_cascadenet.pt	Disease progression transformer
4	Neural ODE	62K	0.772	models/best_neural_ode.pt	Learned differential equation of disease
5	Drug Effect Net	159K	—	models/best_drug_effect_net.pt	Drug repurposing (ARBs, Metformin, Gabapentin)
6	BrainCascadeGNN	125K	0.728	models/best_brain_gnn.pt	Brain GNN that learned Braak staging
7	MultiScaleAlzheimerNet	441K	0.931	models/best_multiscale_alzheimer.pt	First multi-scale biological fusion

Data Sources Used

• ADNI: 15,834 visits, 3,788 patients, 77K medication records
• Allen Brain Atlas: 6 donor brains, 50 AD genes x 68 regions
• Hansen PET Receptors: 41 neurotransmitter maps (5-HT, DA, GABA, mGluR, CB1, opioid...)
• BrainSpan: Developmental transcriptomics (16 pcw to 40 years)
• Bellenguez 2022 GWAS: 20M SNPs, 5,565 genome-wide significant (111K AD cases)
• GEO Postmortem: 6 studies (~1,243 AD/control brains)
• STRING-DB: 42 AD gene protein interaction network
• ChEMBL: BACE1/GSK3B/AChE drug-target activities
• HCP Connectivity: 68x68 structural connectome (ENIGMA)
• Hansen 8-Modal: Gene co-expression, metabolic, haemodynamic, electrophysiological, receptor, laminar, temporal
• Braak Staging: Ground truth tau propagation (stages 1-6)
• Gene Ontology: 8 AD-relevant pathways
• ClinicalTrials.gov: AD trials database

Key Findings

1. Blood-only matches full model (AUC 0.733 vs 0.727) — publishable for screening
2. Temporal change is king — visit-to-visit change (AUC 0.958) far outperforms single-visit
3. BrainCascadeGNN independently learned Braak staging — hippocampus > amygdala > thalamus
4. Receptor density is most informative scale (25.7%) in MultiScale model
5. Drug candidates: ARBs (MMSE +0.50/yr), Metformin (+1.21/yr), PPI+Trazodone synergy

Folder Structure

alzheimer-research-complete/
  models/           — 7 trained .pt model weights
  results/          — JSON results for each model
  scripts/          — All training scripts (self-contained)
  data/
    connectivity/   — HCP + Hansen brain connectivity matrices
    brain_labels/   — 82 brain region labels
    adni_merged_dataset.pkl — Processed ADNI data

Raw Data (4.7 GB)

Raw neuroscience data is in ~/Documents/alzheimer-cascadenet/raw_data/:

• allen_brain/ — 6 donor microarray zips (1.5 GB)
• gwas_ad/ — Bellenguez + Wightman GWAS (1.5 GB)
• neuromaps/hansen_receptors/ — PET receptor NIfTI images (880 MB)
• geo_ad/ — 6 GEO series matrices (278 MB)
• brainspan/ — Developmental expression (264 MB)
• neurosynth/ — 14K fMRI studies (207 MB)
• protein_structures/ — PDB + FASTA for AD targets
• string_db/ — Protein interaction network
• drug_targets/ — ChEMBL activities
• gene_ontology/ — AD pathway genes

How to Reproduce

# 1. Prepare ADNI data (requires RDA files in working directory)
python scripts/prepare_adni_data.py

# 2. Train any model
python scripts/train_cascadenet.py
python scripts/train_blood_predictor.py
python scripts/train_temporal_cascadenet.py
python scripts/train_drug_discovery.py
python scripts/train_brain_gnn.py
python scripts/train_multiscale_alzheimer.py

Author

Satyawan Singh — Infonova Solutions, Leicester, UK Built with Claude (100% AI-assisted research) Date: 2026-04-05