biosynthai 0.1.0

BioSynthAI is a multi-agent platform for neuroscience and biomedical knowledge synthesis.

BioSynthAI

BioSynthAI is a multi-agent platform for neuroscience and biomedical knowledge synthesis.

It integrates diverse biomedical databases and literature sources to automatically retrieve, analyze, and synthesize scientific knowledge across domains. BioSynthAI aims to support research, discovery, and hypothesis generation by combining structured data, publications, and biological knowledge into unified insights.

Key Features:

• Multi-agent architecture for modular and scalable analysis
• Live integration with biomedical databases (PubMed, Gene, UniProt, STRING, and more)
• Automated literature retrieval and synthesis
• Knowledge graph construction and enrichment
• Support for neuroscience and broader biomedical research workflows
• Local caching and conversation-aware memory

Note: Still in active development.

Databases integrated

Database	What it provides
PubMed / PMC	Literature + full text
NCBI Gene	Gene info, GO terms, RefSeq
NLM MeSH API	Dynamic MeSH term resolution
bioRxiv	Recent preprints
STRING	Protein interactions + pathway enrichment
UniProt	Protein function, domains, diseases
Open Targets	Drug-target-disease associations
DANDI Archive	Neurophysiology datasets (NWB)
Allen Brain Atlas	Brain region gene expression
NeuroMorpho.Org	Neuron morphology reconstructions
ClinVar	Variant counts

Installation

From PyPI

pip install biosynthai

From source (development)

git clone https://github.com/tekrajchhetri/biosynthai.git
cd biosynthai
pip install -e .

Usage

CLI (after pip install)

biosynthai "What does BDNF do in the hippocampus?"

# Pass API key directly (no .env needed)
biosynthai "BDNF hippocampus" --api-key sk-or-v1-...

# With NCBI key for higher PubMed rate limits
biosynthai "BDNF hippocampus" --api-key sk-or-v1-... --ncbi-key abc123

# Specify model
biosynthai "SNCA in Parkinson's disease" --model google/gemini-2.5-flash

# More results, deeper related-article expansion
biosynthai "hippocampus theta oscillations" --max-results 20 --depth 2

# Skip some databases
biosynthai "MAPT tau aggregation" --no-dandi --no-neuromorpho

# Export specific format
biosynthai "dopamine reward circuits" --format json --output results/

# Version
biosynthai --version

You can also run via python -m biosynthai "..." or python main.py "..." from the repo root.

Python API

import asyncio
from biosynthai import BioSynthesizer, BioSearchQuery

synth = BioSynthesizer(api_key="sk-or-v1-...")
query = BioSearchQuery(question="What does BDNF do in the hippocampus?")
result = await synth.run(query)

print(result.synthesis_text)
print(f"Articles: {len(result.articles)}")
print(f"DANDI datasets: {len(result.dandi_datasets)}")

Environment variables

Variable	Required	Description
OPENROUTER_API_KEY	Yes	Get one at openrouter.ai/keys
NCBI_API_KEY	No	Raises PubMed rate limit 3 → 10 req/s

All keys can also be passed via CLI flags (--api-key, --ncbi-key), which take priority over environment variables.

Changing the model

Any model available on OpenRouter works:

biosynthai "BDNF" --model anthropic/claude-sonnet-4       # default
biosynthai "BDNF" --model anthropic/claude-haiku-4        # faster/cheaper
biosynthai "BDNF" --model google/gemini-2.5-pro
biosynthai "BDNF" --model openai/gpt-4o
biosynthai "BDNF" --model deepseek/deepseek-chat

Package structure

biosynthai/
├── __init__.py      # Public API
├── models.py        # Pydantic v2 models
├── agents.py        # pydantic-ai agents (search plan, synthesis, profile)
├── clients.py       # All database clients (PubMed, STRING, DANDI …)
├── pipeline.py      # BioSynthesizer orchestrator
├── evidence.py      # Heuristic evidence extraction
├── export.py        # Markdown / JSON / BibTeX export
├── main.py          # CLI entry point (biosynthai command)
└── __main__.py      # python -m biosynthai support

pydantic-ai agents

Agent	Output type	Purpose
search_plan_agent	BioSearchPlan	Generates structured search strategy
synthesis_agent	str	Grounded narrative synthesis
evidence_extraction_agent	BatchEvidenceExtraction	Extracts evidence spans
friendly_profile_agent	FriendlyProfile	Human-friendly gene profile

All agents use defer_model_check=True so the model is injected at runtime, making FastAPI integration straightforward.

License

Apache 2.0 — © 2026 Tek Raj Chhetri