globalnebula/NM-GMED-1

Model Card for NutriGNN — Drug–Food Pharmacological Mapping LLM

NutriGNN is a multimodal, safety-aware biomedical reasoning model fine-tuned on the Meta-Llama-3-8B base using Parameter-Efficient Fine-Tuning (PEFT) with LoRA adapters.
It functions as an explainable AI assistant for drug–food interaction analysis, personalized nutritional recommendation, and biomedical report generation, grounded in a heterogeneous knowledge graph and graph neural embeddings.

---

Model Details

Model Description

NutriGNN integrates Graph Neural Networks (GNNs) with a Large Language Model reasoning layer to align pharmaceutical drug effects with nutraceutical and food-derived compounds.
The system performs cross-domain embedding alignment, safety constraint filtering (allergies, diseases, contraindications), and natural-language explanation/recipe generation.

• Developed by: S Kunal Achintya Reddy
• Model Type: Biomedical reasoning and recommendation LLM
• Languages: English
• License: MIT
• Fine-tuned from model: meta-llama/Meta-Llama-3-8B
• Frameworks: PyTorch, PEFT, Hugging Face Transformers, PyTorch Geometric
• Version: 1.0 (2026)

NutriGNN is designed as a decision-support system, not a clinical diagnostic tool.

---

Uses

Direct Use

NutriGNN can be used directly for:

• Drug → food compound mapping and analog discovery
• Personalized dietary recommendations with safety filters
• Contraindication and side-effect explanation generation
• Biomedical recipe generation and structured reports
• Conversational health copilots and educational assistants

Downstream Use

• Domain adaptation for clinical nutrition or pharmacy analytics
• Integration into healthcare dashboards or wellness platforms
• Embedding as an explainability layer in biomedical RAG systems

Out-of-Scope Use

• Automated medical diagnosis or treatment decisions
• Emergency or critical care decision-making
• Processing personal health data without consent
• Replacing licensed healthcare professionals

---

Bias, Risks, and Limitations

• Outputs depend on the quality and coverage of biomedical datasets used during training.
• The model may underperform in rare drug–food interactions or niche medical domains.
• Not a certified medical advisor; recommendations require professional validation.
• Language and regional dietary diversity may affect accuracy.

Recommendations

Users and developers should:

• Treat NutriGNN as decision support, not authoritative medical advice.
• Enable safety filters (allergy and contraindication modules) during deployment.
• Periodically retrain or fine-tune with updated biomedical datasets.
• Combine with human oversight in clinical or regulated environments.

---

Training Details

Training Data

NutriGNN was trained on a curated biomedical corpus composed of:

• Public pharmacological datasets (drug–target relations)
• Nutraceutical and food compound datasets
• Biomedical literature summaries and ontology mappings
• Synthetic reasoning dialogues and recipe generation prompts
• Structured knowledge graph triples for cross-domain alignment

All data used was publicly available, synthetic, or anonymized.

---

Training Procedure

Preprocessing

• Data normalized into structured triplets: context → biochemical reasoning → safe recommendation
• Ontology mapping and duplicate removal
• Contraindication and allergy tags injected as control tokens
• Recipe and explanation prompts standardized into instruction format

---

Training Hyperparameters

• Base model: Meta-Llama-3-8B
• Fine-tuning: LoRA (r=32, alpha=16, dropout=0.05)
• Batch size: 32–64
• Learning rate: 2e-4
• Optimizer: AdamW
• Precision: bf16 mixed precision
• Epochs: 4–6
• Context length: 4096 tokens
• Training hardware: NVIDIA A100 GPU (80GB)

---

Speeds, Sizes, Times

• Trainable parameters: ~110M
• Adapter checkpoint size: ~2–3 GB
• Average training throughput: ~380 tokens/sec
• Total training time: ~11.8 GPU hours
• Estimated training cost: $21.63 USD
  (Equivalent to ~11–12 hours on a single A100 GPU instance at ~$1.8–2.0/hr spot pricing)

---

Evaluation Metrics

Metric	Purpose
Precision@K	Food analog retrieval quality
AUROC / AUPRC	Side-effect and target prediction
Cosine Similarity Threshold	Pharmacological alignment validity
Acceptance Rate	Recipe biological coherence
Zero-Shot Accuracy	Generalization to unseen drugs
Human Evaluation	Interpretability and clarity

---

System Architecture


Ethical and Safety Considerations

NutriGNN includes safety-aware reasoning modules, but:

• It does not replace medical professionals.
• It should not be used for regulated clinical decisions without validation.
• Developers should implement consent, logging, and explainability layers.

---

Intended Deployment Contexts

• Research prototypes and academic studies
• Educational healthcare tools
• Nutritional planning assistants
• Biomedical RAG or graph-augmented reasoning systems

Not recommended for emergency or high-risk medical scenarios.

---

Citation

If you use this model in academic work:

@misc{nutrignn2026, title={NutriGNN: Graph–Language Modeling for Drug–Food Bioactivity Alignment}, author={S. Kunal Achintya Reddy}, year={2026}, note={Hugging Face Model Repository} }

---

License

MIT License — free for research and commercial use with attribution.

---

Final Note

NutriGNN bridges computational pharmacology and nutritional intelligence by combining graph-based reasoning with generative language models, enabling safer and more interpretable biomedical decision-support systems.