Babajaan/edge-vcf-pan-saudi

🧬 Pan-Genomic Saudi Edge-VCF

A universal, 100% offline, mobile-native clinical variant interpreter for the Saudi population.
Built by BABA AI Solutions using Teacher-Student Distillation with the PAVS 2026 Framework.

Architecture

┌─────────────────────────────────────────────────────────────┐
│  NTv3_650M_post (Teacher, 650M params)                      │
│  InstaDeepAI Nucleotide Transformer v3                      │
│  → Universal Genomic Embeddings (1536-d)                    │
└──────────────────┬──────────────────────────────────────────┘
                   │ TimeKD Distillation
                   │ (KL-Divergence + SmoothL1 Feature Alignment)
                   ▼
┌─────────────────────────────────────────────────────────────┐
│  Edge-RGF-Pan (Student, 76,141 params < 100k budget)        │
│  ┌─────────────┐  ┌─────────┐  ┌──────────────────────┐    │
│  │ Input Proj   │→│ 1D-CNN  │→│ Bi-GRU (bidirect.)    │    │
│  │ 1536→32     │  │ 24ch    │  │ hidden=28, out=56     │    │
│  └─────────────┘  └─────────┘  └──────────┬───────────┘    │
│                                            │                 │
│  ┌────────────────────────────────────────┐│                 │
│  │ FiLM Conditioning Layer                ││                 │
│  │ context = [HPO_Emb(8), Saudi_AF(1),   │├→ Classifier    │
│  │            Consanguinity_F(1)]         ││   → 5 classes   │
│  │ output = (1 + γ) · h + β              ││                 │
│  └────────────────────────────────────────┘│                 │
└─────────────────────────────────────────────────────────────┘

Parameter Budget

Component	Parameters	Share
Input Projection (1536→32)	49,184	64.6%
1D-CNN (24 channels)	4,664	6.1%
Bi-GRU (hidden=28)	10,416	13.7%
HPO Embedder	512	0.7%
FiLM Generator	2,080	2.7%
Classifier	1,989	2.6%
Feature Projection	7,296	9.6%
TOTAL	76,141	< 100k ✓

Package Structure

edge_vcf_pan/
├── __init__.py        # Package metadata
├── config.py          # Saudi allele frequencies, HPO ontology, PAVS 2026 KG
├── teacher.py         # NTv3_650M_post wrapper (+ simulation mode)
├── student.py         # Edge-RGF-Pan: 1D-CNN + Bi-GRU + FiLM (<100k params)
├── distillation.py    # TimeKD loss + training loop
├── agent.py           # smolagents v1.24.0 CodeAgent + 3 tools
├── export.py          # ONNX + INT8 TFLite export
└── main.py            # End-to-end demo

exported_models/
├── edge_rgf_pan.onnx       # ONNX model (77 KB)
└── edge_rgf_pan_int8.tflite # INT8 quantized TFLite (108 KB)

Clinical Agent Tools

Tool	Type	Function
query_pan_saudi_db	Tool subclass	Lazy-loads Edge-RGF-Pan, queries PAVS 2026 knowledge graph
hpo_reasoner	@tool	Maps raw symptoms → HPO IDs → ranked candidate genes
clinical_reporter	@tool	ACMG Class 1-5 pathogenicity classification with Saudi context

Verified Clinical Cases

Case A: Alport Syndrome (Renal)

• Symptoms: "blood in urine, hearing loss"
• Consanguinity: first_cousin (F=0.0625)
• HPO Mapping: HP:0000790 (Hematuria), HP:0012587 (Macroscopic hematuria), HP:0000407 (Hearing impairment)
• Top Gene: COL4A3 → c.765delC (50x Saudi enrichment)
• Classification: Differential renal panel triggered

Case B: Retinal Dystrophy (Vision)

• Symptoms: "retinal spots, night blindness"
• Consanguinity: second_cousin (F=0.015625)
• HPO Mapping: HP:0000580 (Pigmentary retinopathy), HP:0000662 (Nyctalopia)
• Top Gene: RPE65 → c.271C>T (15x Saudi enrichment)
• Classification: Differential retinal panel triggered

✓ Same model, different phenotype contexts → different clinical classifications

Quick Start

pip install torch smolagents onnx onnxscript
python -m edge_vcf_pan.main

License & Disclaimer

Research/clinical decision support only. All findings must be validated by a certified clinical geneticist. Not a medical device.

Built by BABA AI Solutions — Edge-VCF-Pan v1.0