Knowledge Graphs are an increasingly relevant approach to store detailed knowledge in many domains. Recent advances in NLP allow to enrich Knowledge Graphs through automated analysis of large volumes of literature, reducing a lot the efforts in traditional manual information capturing. In our presentation we report the approach taken in a project with partner Fraunhofer SCAI in the life sciences where a knowledge graph organising detailed facts about psychiatric diseases has been computed. Information of cause-effect relations between proteins, genes, drugs and diseases has been encoded in the BEL (Biological Expression Language) and imported into a Graph database to approach an indication-wide Knowledge Graph for the selected therapeutic area. Ultimately, updating the graph will amount to just rerunning the analysis on the newly published literature.

1. “Automatic relation extraction
from biomedical content for
knowledge graph creation and
maintenance”
Stefan Geißler - Kairntech
“AI SDV”
Vienna, Oct 11, 2022
info@kairntech.com

2. 1
2
3
French/German NLP/AI specialist created in December 2018
Grenoble, France (headquarter), Paris and Heidelberg, Germany
Selected customers: Boehringer-Ingelheim, SCAI-Fraunhofer, AFP, French
Government, Groupe Revue Fiduciaire,…
Introducing Kairntech
4 Mission: Making NLP / Machine Learning accessible for domain experts (“No code /
Low code platform”)
Stefan Geißler
Co-founder
Kairntech
Heidelberg

3. Project Intro & Background
• Spring 2021: Fraunhofer SCAI: “We are looking for an automatic
procedure to extract information about entities and their relations from a
larger corpus on psychiatric disorders and render that in the BEL language.
Can you do that?”
“Phosphorylation of glycogen synthase kinase 3beta at Threonine,
668 increases the degradation of Amyloid precursor protein.”
BEL (Biological Expression Language) captures
and encodes biological relations
Copied from Martin Hofmann-Apitius

4. The challenge: updating knowledge graphs
Number of articles
Existing knowledge graphs outdated quickly; update is
time-consuming and challenging
Data volumes are large –
but not infinite
Knowledge is complex -
but can be formalized
Hypotheses:
Publications on Pubmed on “TAU phosphorylation”
Copied from Martin Hofmann-Apitius

5. Assessing the “Pharmacome”
Previous efforts at SCAI: Manual creation of KGs on
genes, proteins, drugs on Neurogenerative diseases
Graphs often involve thousands of
genes, proteins and their interactions
On Alzheimer alone, SCAI
built 24 specific subgraphs
Copied from Martin Hofmann-Apitius
Main goal: Find information for
drug repurposing: Which
drugable, known substances
have a therapeutic effect on the
investigated diseases?

6. Relevant but hard to find: Indirect relations
“A interacts with B which stimulates C which blocks D …”
Approach: Harvest interactions automatically and introduce them
into machine-readable large indication-wide Knowledge Graphs
Chains of interactions over
multiple entities are notoriously
hard to find, often unknown even
to individuals who are aware of
the specific isolated steps.
Copied from Martin Hofmann-Apitius

7. Towards Automatic Knowledge Graph Creation:
Relation Extraction
Knowledge Graphs:
• Network of entities and the relations between
them
• Facilitates investigating many relevant questions
(compared to trad. Relational DBs)
• E.g. in bio-sciences: pathways (“A stimulates B
which blocks C which increases D…”)
• Growing adoption in the industry
• Many intuitive data management systems
Relation Extraction:
• Challenging NLP task: Identify relations from
natural language text
• Co-Occurrence: Relation holds when entities
occur together in text
• Benefits: Fast, simple
• Drawbacks: overgeneration, no relation
type
• Hand-crafted rules: Apply complex grammar
• Benefits: Can be very accurate, can detect
types of relations
• Drawbacks: complex process of rule
building. How to model relations beyond
sentence-level?
• Machine-learning: Training on annotated text
• Benefits: highest quality
• Drawback: Data & Computation intensive

8. Kairntech Studio (build a pipeline)
TXT, PDF,
Word, HTML,
XML, JSON…
Import
documents
Create a training
dataset
Explore, label text with
manual or assisted text
annotation tools
Create and compare
learning models
Create NLP pipelines
Combining models, client taxonomy,
Knowledge graphs, technical
components…
Using built-in and
state-of-the-art
algorithms.
Usage via No-code &
easy-to-use web
application
or
via RestAPI

9. Kairntech Server (run the pipeline in production)
REST API
Link entities to
WikiData taxonomies
Custom housekeeping
to improve
recognition quality:
Filtering
Refine entity
extraction
NLP Pipeline
Knowledge
applications
• Large-scale entity
recognition
• Multi-topic
• Multi-lingual
• Constantly updated
• Linked to background
information
Compute add’tl
entity properties
Proteins receive label
about protein
modification (ex:
phosphorylation)
Build relations
between entities
Format results
to output
Apply custom-
trained relation
extraction model
Return results in
BEL format for
integration in
Knowledge Graph
environment

10. Entity Linking & Namespaces
Entities are not just strings, they need to be associated with the real-world objects they
refer to.
A subset of namespaces is used:
⚫ HGNC (Hugo Gene Nomenclature Committee) for proteins
⚫ MeSH (Medical Subject Headings) for pathologies
⚫ CheBI (Chemical Entities of Biological Interest) for drug/chemical abundances
⚫ GO (Gene Ontology) for biological processes
Recognized entities are
⚫ Disambiguated (“cancer”: animal or disease?)
⚫ Scored (prominence/weight of the concept at in this context?)
⚫ Normalized (synonyms → preferred form, ex: “NIDDM” → “Diabetes Mellitus Type 2”)
⚫ Linked: Entities are linked to world-knowledge

11. NLP Pipeline for Entity Recognition & Linking: Example!

12. Relationship extraction
• Trained on a relation training data set
• The table (right) lists the most important
relations addressed in the project
• The model determines whether between any
pair of entities a relation holds or not
(“NoRelation”) and if yes, which one
• Quality decreases with smaller numbers of
training examples per relationship type (as
expected).
• Manual inspection of (a sample of) the results
from Kairntech by SCAI experts: ~73% of the
relations returned by Kairntech are valid
RELATION NoRelation
RELATION increases
RELATION decreases
RELATION regulates
RELATION positiveCorrelation
RELATION association
RELATION negativeCorrelation

13. Complete NLP Pipeline with Relationship Extraction: Example!

14. Results displayed as a Knowledge Graph: Example!
“Which proteins are known to be positively correlated both
with Schizophrenia as well as Bipolar Disorders?”

15. “We now have relation extraction accuracies of
between 70% and 80%. Not long ago you had to
be happy to get that for just entity extraction.
Having this for relations now is outstanding.”
Martin Hoffmann-Apitius, SCAI
Link entities to
taxonomies
Refine entity
extraction
Kairntech Natural Language Processing Pipeline
Create relations
between entities
BEL
output
Automated relationship extraction

16. Ongoing work
• Speed up the analysis by parallelizing the process: Make use of SCAI’s high
performance computing cluster, leverage SCAI expertise in parallelizing
complex software processes
• Investigate extension to other therapeutic areas
• Define joint approach & offering for industry use cases on
• computing topic specific knowledge graphs
• Updating/extending knowledge graphs
• Expand approach to cover large chunks (all?) of Medline?
• Joint (Kairntech&SCAI) communication efforts: Webinars, publications
• Cf. www.biorxiv.org/content/10.1101/2022.03.07.483233v1

17. Conclusion / findings
• Kairntech off-the-shelf entity extraction performs well even in this
highly specific subdomain
• Kairntech notion of processing pipelines allows to define
sophisticated processing chains (here: entity recognition, application
of specific custom model (ModType), relation extraction, output
encoding into BEL syntax)
• Relation results assessed and declared useful by SCAI experts after
detailed manual analysis
• Relation extraction from large literature corpus to feed Knowledge
Graphs is feasible

18. Thank you for your
attention!
info@kairntech.com