BioMedOS

Local-first multi-agent biomedical operating system with knowledge graph reasoning, PubMed RAG, graph ML, and clinical decision support. Your private biomedical AI assistant.

BioMedOS combines a biomedical knowledge graph, literature retrieval, graph machine learning, clinical tooling, and a 12-agent orchestration layer into one local stack. It is designed to run with Ollama, ChromaDB, and Python on your own machine without cloud model dependencies.

Why BioMedOS?

• Local-first by design: Ollama, ChromaDB, and the core graph stack run on local infrastructure with no required cloud LLM APIs.
• KG + RAG hybrid reasoning: retrieval merges BM25, dense search, and knowledge graph context so answers are grounded in both literature and structure.
• 12 specialized agents: routing, literature synthesis, graph traversal, link prediction, repurposing, clinical reasoning, review writing, and verification are all separated into focused modules.
• Clinical decision support built in: DDI checks, phenotype mapping, differential diagnosis, contraindication review, and evidence grading are part of the same platform.

Architecture

                         +----------------------------------+
                         |           Web UI / API           |
                         |  SPA + FastAPI + WebSocket chat  |
                         +----------------+-----------------+
                                          |
                           +--------------v--------------+
                           |       Agent Orchestration    |
                           | Router -> Specialists ->     |
                           | Sentinel -> Aggregation      |
                           +--------------+--------------+
                                          |
        +------------------------+--------+---------+------------------------+
        |                        |                  |                        |
+-------v--------+     +---------v---------+ +------v------+     +-----------v-----------+
|  Data Sources  |     | Knowledge Graph   | | RAG Engine  |     | Clinical Tooling      |
| PubMed, OT,    | --> | NetworkX schema   | | BM25+dense  | --> | DDI, phenotypes,      |
| ChEMBL, HPO... |     | queries, stats,   | | KG context, |     | diagnosis, evidence   |
|                |     | graph ML export   | | reranking   |     | grading               |
+----------------+     +---------+---------+ +------+------+     +-----------+-----------+
                                          |                  |
                                 +--------v------------------v--------+
                                 |     Graph ML + Local Ollama        |
                                 | GraphSAGE, R-GCN, Node2Vec, LLMs   |
                                 +------------------------------------+

Quick Start

python -m venv .venv
. .venv/Scripts/Activate.ps1
pip install -e .[dev]
python scripts/demo.py
python scripts/run_local_api.py --port 8010

Alternative entrypoints:

docker compose up -d
python scripts/build_graph.py --genes EGFR TP53 BRCA1 ALK --sources open_targets string_db pubmed
python scripts/train_models.py --model graphsage --epochs 50 --edge-type gene_associated_with_disease
uvicorn biomedos.api.app:app --reload

Data Sources

API	Purpose
PubMed E-utilities	Search and fetch abstracts for literature grounding
PubTator	Biomedical annotations for literature enrichment
Open Targets	Gene-disease associations, tractability, and drug evidence
ChEMBL	Drug metadata, mechanisms, bioactivities, and ADMET signals
UniProt	Protein metadata and gene-to-protein lookup
STRING DB	Protein-protein interaction context
RxNorm	Drug normalization, interactions, and NDC mappings
Human Phenotype Ontology	Phenotype search, disease links, and phenotype similarity
OpenFDA	Adverse events, labels, contraindications, and recalls
DisGeNET	Gene-disease association evidence
ClinicalTrials.gov v2	Trial metadata for translational and clinical workflows

Knowledge Graph Schema

Node types: Gene, Protein, Disease, Drug, Compound, Pathway, Phenotype, CellType, Tissue, SideEffect, ClinicalTrial, Publication

Edge types: gene_associated_with_disease, gene_in_pathway, gene_interacts_with_gene, protein_interacts_with_protein, drug_targets_gene, drug_treats_disease, drug_interacts_with_drug, drug_causes_side_effect, compound_binds_target, disease_has_phenotype, disease_involves_pathway, gene_expressed_in_tissue, pathway_crosstalks_with, publication_mentions_gene, publication_mentions_disease, drug_contraindicated_for, gene_associated_with_phenotype, trial_investigates_drug

Agents

Agent	Default model	Role
Router	llama3.2:3b	Classifies requests and decomposes complex workflows
Literature	qwen2.5:14b	Runs PubMed-grounded retrieval and cited synthesis
Graph Explorer	qwen2.5:14b	Traverses paths, neighborhoods, and local subgraphs
Link Predictor	local GNN + qwen2.5:14b	Scores novel graph links and explains predictions
Drug Repurposer	local GNN + qwen2.5:14b	Finds cross-disease drug opportunities with confidence labels
Geneticist	qwen2.5:14b	Builds gene-centric reports with network and druggability context
Pharmacologist	qwen2.5:14b	Reviews DDI, PK, ADMET, safety, and contraindications
Clinician	qwen2.5:14b	Produces differential diagnoses from symptoms and phenotypes
Pathway Analyst	qwen2.5:14b	Runs enrichment and pathway crosstalk analysis
Hypothesis Generator	qwen2.5:14b	Surfaces structural holes and ranked mechanistic hypotheses
Review Writer	qwen2.5:14b	Drafts narrative reviews with citations and self-critique
Sentinel	phi4:14b	Verifies claims, citations, and hallucination risk

Clinical Decision Support

• Drug-drug interaction checking via RxNorm and OpenFDA
• Symptom to HPO term mapping with phenotype matching
• Differential diagnosis combining phenotype evidence, KG support, and literature
• Contraindication review from local label parsing
• Evidence grading with GRADE-style classifications

Graph ML

Model	Purpose	Status	Metrics
GraphSAGE	Heterogeneous link prediction baseline	Implemented	See RESULTS.md
R-GCN	Multi-relational reasoning with relation-specific weights	Implemented	Placeholder
Node2Vec	CPU-friendly embedding baseline	Implemented	Placeholder

RAG Pipeline

Query
  -> BM25 sparse retrieval
  -> Dense vector retrieval
  -> Knowledge graph context extraction
  -> Reciprocal rank fusion
  -> Cross-encoder reranking
  -> Grounded answer generation with citations

Comparison

System	Local-first	Knowledge graph reasoning	Multi-agent orchestration	Clinical tooling
BioMedOS	Yes	Yes	Yes	Yes
KG-RAG	Partial	Yes	No	No
PaperQA2	Partial	No	Limited	No
STELLA	Partial	Limited	No	No
DrugAgent	Partial	Drug-focused	Yes	Limited
MedRAG	Partial	Limited	No	Limited

Hardware

• 32 GB RAM: practical baseline for 8B-class models and CPU-first experimentation
• 64 GB RAM: recommended for the default 14B reasoning and verification stack
• 128 GB RAM: suitable for larger local models, heavier indexing, and broader workflows

Results

Benchmark placeholders and experiment templates live in RESULTS.md.

References

1. PubMed E-utilities documentation
2. Open Targets Platform
3. ChEMBL Web Services
4. STRING database
5. Human Phenotype Ontology
6. ClinicalTrials.gov API v2
7. PyTorch Geometric documentation
8. LangGraph documentation

License

MIT. See LICENSE.