AI-powered biomedical evidence intelligence platform for literature synthesis, knowledge graph generation, and evidence-aware scientific reasoning.
BioGraph Intelligence transforms biomedical literature into structured scientific knowledge.
Instead of only retrieving PubMed papers, the platform automatically retrieves relevant biomedical literature, extracts structured scientific claims using a local LLM, clusters related evidence into themes, scores evidence strength, detects conflicting findings, identifies research gaps, generates an AI research brief, builds an interactive biomedical knowledge graph, and enables evidence-grounded question answering through an AI Copilot.
The goal is to reduce literature review from hours of manual reading to minutes of AI-assisted evidence synthesis.
Biomedical literature is growing faster than researchers can manually review. Existing search tools retrieve papers, but they often do not synthesize evidence across studies, identify contradictions, or highlight research gaps.
BioGraph Intelligence was built to explore how AI systems can convert unstructured biomedical publications into structured, explainable, decision-ready evidence.
- PubMed search integration
- PMID, title, journal, author, DOI, and abstract extraction
- Query-driven biomedical evidence collection
Each paper is converted into a structured biomedical claim.
Example:
Subject: SGLT2 inhibitors
Relationship: REDUCES_RISK_OF
Object: kidney failure progression
Evidence Type: cohort study
Confidence: medium
Stance: support
PMID: 12345678
Related claims are grouped into higher-level biomedical evidence themes.
Evidence themes are ranked based on current support signals such as paper count, confidence, and evidence type.
The system identifies whether claims are supportive, contradictory, or inconclusive.
The system identifies limitations such as:
- limited evidence coverage
- lack of longitudinal validation
- lack of clinical trial evidence
- unclear population-level generalizability
The platform generates a structured executive research brief covering:
- overall evidence
- key findings
- current consensus
- important limitations
- future research directions
Extracted claims are visualized as a biomedical knowledge graph.
Subject → Relationship → Object
Users can ask follow-up questions grounded in the extracted evidence from the current analysis run.
Biomedical Research Question
↓
PubMed Retrieval
↓
Paper Objects
↓
Local LLM Claim Extraction
↓
Structured Claims
↓
Evidence Theme Aggregation
↓
Evidence Scoring
↓
Conflict Detection
↓
Research Gap Detection
↓
AI Research Brief
↓
Knowledge Graph
↓
Evidence Copilot
KnowledgeGraph/
├── app/
│ ├── llm/
│ │ ├── base.py
│ │ ├── factory.py
│ │ └── ollama.py
│ │
│ ├── models/
│ │ ├── paper.py
│ │ ├── claim.py
│ │ ├── evidence_theme.py
│ │ ├── research_brief.py
│ │ ├── research_gap.py
│ │ └── conflict.py
│ │
│ ├── services/
│ │ ├── pubmed.py
│ │ ├── extractor.py
│ │ ├── aggregator.py
│ │ ├── scorer.py
│ │ ├── research_brief.py
│ │ ├── gap_detector.py
│ │ ├── conflict_detector.py
│ │ ├── copilot.py
│ │ └── graph_builder.py
│ │
│ └── main.py
│
├── data/
├── docs/
├── requirements.txt
└── README.md
| Category | Technology |
|---|---|
| Language | Python |
| UI | Streamlit |
| Local LLM | Ollama |
| Literature Source | PubMed |
| Graph Visualization | PyVis |
| Data Modeling | Pydantic |
| Data Processing | Pandas |
The project was migrated from a paid API-based LLM workflow to a local Ollama-based LLM workflow.
Benefits:
- no API credit dependency
- local development
- privacy-preserving inference
- reproducible testing
- easier iteration during development
Instead of generating free-text summaries only, the system converts each paper into a structured claim object containing:
- subject
- relationship
- object
- population
- evidence type
- confidence
- stance
- PMID
- publication metadata
This enables downstream aggregation, scoring, conflict detection, graph construction, and copilot reasoning.
Initial extraction used one large prompt for multiple papers. This worked for small runs but became unreliable with larger retrieval sizes because the local LLM ignored the required JSON format.
The pipeline was redesigned to extract claims one paper at a time.
Before:
20 papers → one large prompt → unreliable JSON
After:
Paper 1 → claim
Paper 2 → claim
Paper 3 → claim
...
Merge claims → aggregate evidence
This makes extraction slower but more reliable and scalable.
Problem
The initial system used an external LLM API for extraction and semantic grouping. During development, repeated calls quickly exhausted available credits.
Solution
The LLM layer was refactored to use Ollama locally.
Result
The project can now be developed and tested without API costs.
Problem
Streamlit reruns the entire script whenever a button is clicked. This caused the full literature retrieval and claim extraction pipeline to restart when using the Copilot.
Solution
Implemented st.session_state to persist papers, claims, evidence themes, research gaps, conflict summaries, and research briefs.
Result
Follow-up interactions reuse the previous analysis instead of rerunning the expensive pipeline.
Problem
For larger prompts, the local LLM sometimes returned narrative summaries instead of valid JSON.
Solution
The extraction architecture was changed from batch extraction to paper-level extraction. The parser also includes JSON extraction, validation, and retry logic.
Result
The system became more reliable for larger literature searches.
Example query:
How do GLP-1 receptor agonists compare with SGLT2 inhibitors for renal outcomes?
The system can retrieve relevant studies, extract comparative claims, identify evidence themes, surface research gaps, and generate a knowledge graph of drug-outcome relationships.
Example query:
What is the evidence supporting pTau217 as a blood-based biomarker for Alzheimer's disease?
The system can synthesize biomarker claims, identify evidence strength, and generate an AI research brief.
- Claim extraction currently relies primarily on abstracts rather than full-text papers.
- Evidence scoring is an early prototype and does not yet fully weight sample size, study quality, or journal impact.
- Conflict detection depends on the quality of extracted stance labels.
- Graph visualization is optimized for moderate-sized evidence sets and will require filtering for larger corpora.
- Ontology normalization with MeSH, UMLS, DrugBank, or MONDO is not yet implemented.
- Add evidence scoring v2 with study design and sample size weighting
- Add PDF evidence report export
- Improve graph filtering and node type classification
- Add MeSH/UMLS ontology normalization
- Add temporal evidence tracking
- Add persistent caching by PMID
- Add Neo4j graph database backend
- Add full-text paper ingestion
- Add multi-query evidence comparison
- Add deployment-ready FastAPI backend
This project reinforced that building useful AI systems requires more than connecting an LLM to a user interface. The most important engineering challenges were:
- designing reliable data flow
- converting unstructured text into structured objects
- validating LLM outputs
- managing application state
- separating UI, services, models, and LLM logic
- making evidence explainable rather than opaque
BioGraph Intelligence is an end-to-end applied AI system that combines biomedical NLP, local LLM orchestration, structured data modeling, evidence synthesis, and knowledge graph visualization.