ARCHITECTURAL_BLUEPRINT.md

EpiADR-Net Architectural Blueprint & Tensor Operations

This document outlines the detailed tensor dimensions, layer specifications, and mathematical operations across the EpiADR-Net neural pipeline.

---

📐 Tensor Operations Map

Input SMILES String
       │
       ▼
Atom Feature Extraction ─────────► Node Matrix X  : [N, 5]
Edge Connectivity       ─────────► Edge Index E   : [2, M]
GTEx Tissue Profile     ─────────► Tissue Vector V: [B, 128]
       │
       ▼
Linear Input Projection ─────────► h_0             : [N, 128]
       │
       ▼
4-Layer GAT Backbone    ─────────► h_4             : [N, 128]
                                   Attention α_4   : [M]
       │
       ▼
FiLM Modulation Layer   ─────────► γ, β            : [B, 128] -> [N, 128]
                                   h_conditioned   : [N, 128]
       │
       ▼
Dual Graph Pooling      ─────────► [Mean || Max]   : [B, 256]
       │
       ▼
10-Class Classification ─────────► Logits          : [B, 10]
                                   Probabilities μ : [B, 10]
                                   Uncertainty σ   : [B, 10]

---

🔬 Layer Specifications

Layer Name	Input Shape	Output Shape	Parameters / Operations
Input Linear	[N, 5]	[N, 128]	$W_{\text{in}} \in \mathbb{R}^{5 \times 128}$, Bias
GAT Layer 1-4	[N, 128]	[N, 128]	$W_{\text{gat}} \in \mathbb{R}^{128 \times 128}, a \in \mathbb{R}^{256 \times 1}$, LeakyReLU
Gene Pathway Cross-Attn	[N, 1536]	[N, 1536]	$\text{MultiHeadCrossAttn}(\mathbf{Q}{1536}, \mathbf{K}{1024}, \mathbf{V}_{1024})$
FiLM Residual Update	[N, 1536]	[N, 1536]	$h_i^{\text{cond}} = h_i \odot (1.0 + \gamma_{\text{nodes}}) + \beta_{\text{nodes}}$
Dual Graph Pool	[N, 128]	[B, 256]	MeanPool [B, 128] || MaxPool [B, 128]
Classifier FC1	[B, 256]	[B, 128]	Linear(256->128), ReLU, Dropout(0.2)
Classifier FC2	[B, 128]	[B, 10]	Linear(128->10)