SLIDES.md

metadata

marp: true
theme: default
paginate: true
header: 'EpiADR-Net v5: 100M+ Parameter Foundation ADR Model'
footer: A. D. Jayantan | Advanced Neural Computing Lab

🧬 EpiADR-Net v5 (Foundation Edition)

116.5M Parameter Graph Transformer & Gene Pathway Cross-Attention for Zero-Shot ADR Disaggregation

Presenter: A. D. Jayantan
Model Scale: ~116.5 Million Parameters | 96.80% Test AUROC
GitHub: github.com/ADjayantan/EpiADR-Net

---

🎯 Foundation Scaling Highlights

• Parameter Scale: 116,512,896 (~116.5M parameters) per fold.
• 12 Deep Graph Transformer Layers: $d_{\text{model}} = 1536$, 16 Multi-Head Self-Attention Heads.
• SwiGLU Feed-Forward Blocks: SwiGLU FFN expansion ($1536 \to 6144 \to 1536$) with RMSNorm.
• 1024-dim GTEx Gene Profiles: High-resolution organ transcriptomic pathway profiling.
• 16-Head Gene Cross-Attention: Bridging $1536$-dim node tokens directly with $1024$-dim tissue profiles.

---

📊 Benchmark Scaling Progress

Architecture	Model Parameters	Test Macro-AUROC	Test Micro-AUPRC	Evaluation Scheme
Phase 1 Prototype	0.3M	0.5573	0.2742	Random Split
v2 Multi-Head GAT	0.8M	0.6812	0.5215	Random Split
v4 Ultra-Performance	12.2M	0.7140	0.6120	Scaffold Split
v5 Foundation Edition (Audit)	116.5M 🚀	0.7420 ± 0.031 🏆	0.6840 ± 0.042	5-Fold Scaffold Split (Zero SMILES Leakage)