4. Novo Nordisk®
Knowledge Graphs LLMs
Pros:
• Structural knowledge
• Accuracy
• Decisiveness
• Interpretability
• Domain-specific knowledge
Pros:
• Generalized “knowledge”
• Superior language processing
capabilities
• Generalizability
• Hallucination
Cons:
• Incomplete
• Lacks language understanding
• Unseen facts
Cons:
• Hallucination
• Implicit knowledge
• Black-box
• Indecisiveness
• Lack domain-specific knowledge
Adapted from [2306.08302] Unifying Large
Language Models and Knowledge Graphs: A
Roadmap (arxiv.org)
5. Novo Nordisk®
Business value impact
1. Improve data search efficiencies
2. Systematize domain knowledge capture from experts
3. Harness user thought processes, i.e. expert and institutional knowhow,
to refine agents
4. Accelerate data to knowledge to insight cycle for users
8. Novo Nordisk®
Drug Disease
Protein Pathway
Protein Pathway
Protein
Mechanism of Action of Drug to Disease
LLM/Graph-
based predicted
link
treats
targets
downregulates
ameliorates
ameliorates
downregulates
downregulates
downregulates
targets
10. Novo Nordisk®
Use Case: Clinical omics data for treatment
signatures and MoA for drug repurposing
Source: DrugMechDB
Example of a metapath showing the mechanism of action of liraglutide
11. Novo Nordisk®
Modular KGs with KG-aware agentic LLMs
Query_Planner
Research_Assistant
User_Proxy
Summary_writer Reviewer
KG_Assistant
Brainstorm_Expert
KG_Planner
KG Tools
uses
• Get a list of nodes from a description or name
• Count number of nodes
• Get k-shortest paths between a drug and disease node
• Get a list of all possible node types
• Get neighbors of a root node
15. Novo Nordisk®
Working towards fully connected end2end schema
Protocol
CDISC
Controlled
Terminology
NN ontologies
SDTM ADaM Omics
Imaging
Target validation
Digital
Biomarkers
Signatures
Targets
Biomarker development
Drug repurposing
Endpoints
New trials
NN KG schema
RWD
16. Novo Nordisk®
Knowledge Graphs LLMs
Adapted from [2306.08302] Unifying Large
Language Models and Knowledge Graphs: A
Roadmap (arxiv.org)
Agents
Common semantic layer
Communication
Editor's Notes
THZO
In this presentation, we are going to explore how Knowledge Graphs and Language Model Models (LLMs) can work together to unlock the true potential of artificial intelligence and machine learning.
We will begin by exploring the basics of Knowledge Graphs and LLMs, discussing the relationship between the two and how they can be combined to create more powerful and accurate models. Then, we will dive into the many benefits of using Knowledge Graphs to integrate with LLMs, highlighting several real-world use cases. Finally, we will explore practical considerations and best practices for implementing this powerful combination of technologies.