AstraZeneca share their experience of share their experience of building a knowledge graph platform and central service, to power the next generation of insights and analytics at AstraZeneca.
Enhancing Worker Digital Experience: A Hands-on Workshop for Partners
ASTRAZENECA. Knowledge Graphs Powering a Fast-moving Global Life Sciences Organisation.
1. Graph Summit | London | February 22nd, 2024
Knowledge Graphs powering a fast-
moving global life sciences organisation
Our experience building a knowledge
graph platform and service to power the
next generation of insights and analytics at
AstraZeneca
Varun Bhandary
Senior Solutions Architect
Enterprise Data & AI Architecture
IGNITE (AZ)
Antonio Fabregat, PhD
Knowledge Graph Lead
Enterprise Data Office
IGNITE (AZ)
2. Agenda
1. Connected Data ❤️ Lifesciences
2. Our Challenges and Plan 🚏
3. Introducing AZ’s “Knowledge Graph Service" 📣
4. A glimpse into the future of Graphs in AZ 🔭
5. Talking to your Graphs 🗣️🎙️
6. Graphs are Stronger Together ️
2
6. Why Knowledge Graphs at Lifesciences?
6
Integration of Diverse
Data Sources
A unified framework for
connecting heterogeneous
data, enabling researchers and
decision-makers to gain
comprehensive insights across
disparate data silos.
Complexity of Biomedical
Knowledge
Facilitate advanced analytics,
hypothesis generation, and
decision support for drug
discovery, development, and
clinical research.
Semantic Search and
Discovery
Enable semantic search and
discovery by encoding
relationships between entities,
concepts, and attributes in a
graph-based data model
Data-driven Insights and
Decision Making
A powerful foundation for
advanced analytics, machine
learning etc enabling
researchers to uncover
hidden patterns
8. Knowledge Graphs representation alternatives
8
* Adapted from documentation at W3C https://www.w3.org/
Two ways of representing/storing a Knowledge Graph
RDF-star (Resource Description Framework)
Semantic Web: Good for common standards and data exchange
Data model based on 3 parts: subject, predicate and objects
Nodes’ properties added as predicates. Edges with properties are “triple-resources” (like “meta-nodes”)
Storage: “Triple/Quad Stores” Graph Databases
Any type of real-world information, can be represented in a Knowledge Graph
18 nodes (5 instances, 4 classes, 8 literals, 1 triple-resource)
19 relationships (triples)
Knowledge Graph is a way of organizing data & information in the form of a graph
A collection of interlinked concepts, entities, events that represent a network of real-world entities, the relationships between them.
LPG (Labelled-Property Graph)
Good for highly dynamic, transactional use cases
Data organized as nodes, labels, relationships and properties
Both nodes and edges can have properties
Storage: Native Graph Databases
5 nodes (5 ids, 4 Labels, 8 properties)
4 relationships (2 properties)
10. Challenges
10
Decouple & Specialise Integrate & Standardise Abstract & Automate
Use the right tools for the job
Data Lake? Data Warehouse? Graph
Database? LPG? RDF? No-SQL?
Modular Design with Security in
Mind
Build a component-based
architecture with coherent and
practical principles.
Think of data as a product
Push and Pull Vs Serve and Consumer
Make it easy to work with data
across platforms.
Searching and moving data is costly.
Move to an ELT model, leverage
first-party connectors, and
document to promote the most
optimal options.
Standardise
Apply FAIR principles
Document and Promote
Patterns
Data Movement, Loading,
Transformation.
Template and Accelerate
Teams should be able to spend more
time analysing data and deriving
insights than managing infra.
Automate
Leverage IaC, and automation
pipelines to achieve consistent
deployments.
11. The Plan
Data Platform
Unified Data Store
Snowflake
External Tables
Snowflake Internal
Table Storage
Unified Data Compute
Snowflake Virtual
Warehouse
Snowflake
Snowpark
SnowPipe
User Defined
Functions
Unistore
Time-Travel
Data-Lake Compute
SQL Cluster
General Purpose
Cluster
Data Lake Store
Raw Layer
Work Layer
Publish Layer
Glue Hive
Metastore
Knowledge Graph Service
Graph Data Store
LPG Storage
Composite
Utilities
Graph DS Libraries
Cypher / APOC
Graph Compute
Graph Build and
Exploration
Graph Analytics
Machine
Learning Studio
Model Build &
Train
Deploy and
Govern
Graph Exploration
Query Client
Data Browser
Graph Data
Visualization
External Data, RWE &
Partnerships
Structured Data
MDM/RDM, Ontologies,
Vocab., Dictionaries
Semi-Structured
Content & Files
Un-Structured
Content & Files
User Input
Data Acquisition
Data Sources Ingestion &
Integration
IoT &
Streaming
API
Management
Event
Store
Queue
MuleSoft
CDC
Database
API
Streaming
Compute
External Data
Transfer
DDTS
Enterprise Platforms
(i.e. SAP)
Decreasing Volume of Content
Increasing Quality of Content
13. Why Knowledge Graphs? and why a Service?
13
• Data management and analysis
• Overcoming data silos and integration challenges
Growing importance of knowledge graphs
• Hosting and development support for knowledge graphs
• Robust and scalable solutions
• Enhanced data-driven decision-making
Need for efficient and reliable services
• Improved data accessibility and insights
• Streamlined collaboration and innovation
Benefits for businesses and organizations
18. Biology | Market Strategy | Logistics | Environmental targets
18
Biological Insights
Knowledge Graph
Graph machine learning to help scientists
make faster & better drug discovery decisions
Competitive Intelligence
Knowledge Graph
One-stop-shop for competitive intelligence,
transforming a manual system into a rich service
Supply Chain
Knowledge Graph
Insights into the company’s supply chain,
streamlining processes to enhance decision-making
Sustainability
Initiative
Decision-making support system aiming to
reduce the company’s carbon footprint
19. Compounds
19
Compounds Synthesis
& Management
(CSMKG)
Combine several databases
Transforms operational data into business
insights to drive continuous improvements
in storage, logistics and delivery
High Throughput
Screening
(HTSKG)
Contains £M worth of data
Increases the quality and efficiency
of future HTS screens
Compounds
& Fragments
(CFKG)
Creates a view of the chemical space
like a medicinal or computation chemist.
Contains all internal and selected external
libraries and allows users to modify a
search and receive feedback ‘live’
20. PharmaSci
20
Formulation
Knowledge Graph
Pre-clinical formulation design process
Leading to quicker, more effective
scientific developments
Boston Formulation
Knowledge Graph
Improves the understanding of our data
Enhances collaboration by breaking down
silos and connecting disparate data sources
Lipid Nano Particles
Knowledge Graph
Machine learning models
Predicts in-vivo activity from in-vitro
data for intra-cellular drug delivery
and LNP formulation design
21. Talking to your graphs
🗣️🎙️
5
Have you ever thought to
have a graph expert with
you 24/7?
GenAI is here to help!
23. Future Evolutions of the Insights Chat
Knowledge Mesh?
23
Unified Rule, Behavior &
Meta Graph Store
User
User
Knowledge Discovery
Interface
Unified LLM
Integration
(AI Portal)
1
2
3
Domain Specific Knowledge Graphs Domain Specific Knowledge Graphs
Meta Graph Meta Graph Meta Graph
24. Graphs are Stronger
Together ️
Why query federation is a
key to unlocking even more
cross-functional use-cases
6
26. 26
Let’s build bridges to connect “siloes” of interest…
Query federation describes a collection of
features that enable users and systems to
run queries against multiple siloed data
sources without needing to migrate all data
to a unified system.
Federated Queries
are these BRIDGES
27. 27
Let’s build bridges to connect “siloes” of interest…
The diagram shows the resulting subgraph for
the federated query that answers the question
“Find all genes in BIKG linked with a specific disease, and then
all trials in CIKG that are testing drugs targeting those genes”
Biological Insights
Knowledge Graph
Competitive Intelligence
Knowledge Graph
CIKG
28. Acknowledgments
• Aaron Holt
• Nicolas Mervaillie
• Joe Depeau
• Job Maelane
• Yuen Leung Tang
• Jesus Barrasa
• Morgan Senechal
• Lauren Eardley
• Cinthia Willaman
• Taylan Sahin
• Melanie Hardiman
• Daniel Addison
• Delyan Ivanov
• Suzy Jones
• Andriy Nikolov
• Cristina Mihetiu
• Michaël Ughetto
• Karen Roberts
• Wolfgang Klute
• Michael Lainchbury
• Justin Morley
• Andy Stafford-Hughes
• Nikil Kunnappallil
• Anthony Puleo
• Ivan Figueroa
• Koushik Srinivasan
• Nick Iles
• Lena Becciolini
29. Enterprise Data Office | IGNITE
Enterprise Knowledge Graph
Robert Hernandez
Knowledge Engineering
Lead
Sandra Carrasco
Senior Knowledge
Graph Engineer
Antonio Fabregat
Knowledge Graph Lead
Vishal Kumar
DevOps & Data
Engineer
Preetha Mutharasu
Knowledge Graph
Engineer
Ronnie Mubayiwa
Senior DevOps Engineer
Varun Bhandary
Senior Solution Architect
Sree Balasubramanyam
Senior IT Project Manager
Prem Oliver Vincent
Scrum Master
Sangeetha Natarajan
Testing Manager
Miquel Monge
Knowledge Graph
Engineer
Pascual Lorente
Senior Knowledge
Graph Engineer
Santanu Biswas
Senior Datalake Engineer
Tarik Sidi-Mammar
Data Ops Platforms
Service Lead
Lauren Eardley
Enterprise Head of Data
Engineering Services
Notes de l'éditeur
83,000 people across the globe
Main disease areas: Oncology, BioPharma, Rare Diseases
Corporate Approved film for Data and AI at AstraZeneca
Integration of Diverse Data Sources:
Vast amounts of data from various sources (clinical trials, scientific literature, genomic data, patient records, and regulatory documents).
Knowledge graphs provide a unified framework for integrating and connecting heterogeneous data, enabling researchers and decision-makers to gain comprehensive insights across disparate data silos.
Complexity of Biomedical Knowledge:
The field of life sciences is constantly evolving, with new discoveries, insights, and publications emerging rapidly.
Knowledge graphs capture biomedical knowledge's complex relationships and semantics, such as drug-target interactions, disease pathways, genetic associations, and adverse drug reactions.
By organising and representing this knowledge in a structured format, knowledge graphs facilitate advanced analytics, hypothesis generation, and decision support for drug discovery, development, and clinical research.
Semantic Search and Discovery:
Traditional keyword-based search approaches often need help to capture the rich semantics and context of biomedical information.
Knowledge graphs enable semantic search and discovery by encoding relationships between entities, concepts, and attributes in a graph-based data model.
This allows researchers to navigate and explore biomedical knowledge more intuitively and context-awarely, facilitating the identification of relevant insights, hypotheses, and research opportunities.
Data-driven Insights and Decision-Making:
In today's data-driven healthcare landscape, pharmaceutical companies need robust analytics and decision-support tools to extract actionable insights from diverse datasets.
Knowledge graphs are a robust foundation for advanced analytics, machine learning, and predictive modelling, enabling researchers to uncover hidden patterns and leverage the power of
The Data Analytics Funnel
Growing importance of knowledge graphs in data management and analysis.
There is need for an efficient and reliable service to support both hosting and development of knowledge graphs.
AZ is investing in this area to create a robust and scalable service.
We are particularly excited about the future evolutions of our Insights Chat. Our next evolutionary steps in this journey we hope would unlock –
Context switching between domains when interacting with a ChatBot powered by LLMs and underpinned by the different Knowledge Graphs curated and managed by functional teams
The ability to connect across Knowledge Graphs at query time unlocks huge potential with regards to JIT – Just in Time Insights.
Ability to observe, monitor and apply AI and Data governance consistently across projects on LLM-powered RAG applications
Receive real-time user feedback.
We’ve all heard about Data Mesh, but maybe Knowledge Mesh is due to arrive on the Hype Cycle Soon?
When we talk about multiple siloed databases, we could imagine an archipelago. At the first glance, visiting all islands, doesn't seem an easy task!
With the right infrastructure, multiple islands can be connected, and visiting them, suddenly, becomes way easier. Federating queries, across siloed databases, is like building bridges between islands.
This allows running queries against multiple siloed data sources, without needing to migrate all data to a unified system.