Artificial Intelligence: Transforming Pharma Congress

Hilton Boston Dedham Hotel

Charles Room · 4:30 PM

February 21, 2018

Slides at slides.com/dhimmel/dedham

http://www.greenelab.com/

Abstract

Approaches in network medicine have traditionally focused on generating insights from graphs with a single type of node and relationship. However, biology's complexity demands a richer network structure capable of integrating diverse, multi-scale information. Towards this end, we develop hetnets — networks with multiple types of nodes and relationships.

Specifically we created Hetionet — a network of biology, disease, and pharmacology. This resource encodes knowledge from millions of biomedical studies over the last half century. Version 1.0 contains 47,031 nodes of 11 types and 2,250,197 relationships of 24 types. We host a public Neo4j database instance at https://neo4j.het.io allowing users to interact with Hetionet.

In Project Rephetio, we applied Hetionet to predict new uses for existing drugs. Our approach learned the network patterns of connectivity that differentiate treatments from non-treatments, enabling us to predict the probability of treatment for 209,168 compound–disease pairs. These predictions prioritize treatments under investigation by clinical trial.

Going forward, we're investigating more efficient algorithms for feature extraction on hetnets. In addition, we're looking to automate hetnet construction by text mining the literature. The success of hetnets will depend on the availability of openly licensed inputs. As such, I'll briefly discuss data analyses I've performed in hopes of making science more open.

(event details)

How I became intestested in graphs

http://blog.dhimmel.com/friendship-network/

My Facebook friendship network in 2014

Graphs are composed of:

• Nodes
• Relationships

Nodes / relationships have type:

• node types
  (person, course, university)
• relationship types
  (lecturer, institution)

Visualizing Hetionet v1.0

• Hetnet of biology for drug repurposing
   
• ~50 thousand nodes
  11 types (labels)
   
• ~2.25 million relationships
  24 types
   
• integrates 29 public resources
  knowledge from millions of studies

Hetionet v1.0

• Nodes
  • standardized vocabularies
  • stable, unambiguous identifiers
     
• Relationships:
  • Omics scale required
  • Literature mining
  • High throughput experimental technologies
  • Avoid manual mapping
     
• Versioned data dependencies

Constructing Hetionet v1.0

What's the best software for storing and querying hetnets?

neo4j.het.io

• Customized Docker image
• Digital Ocean droplet
• SSL from Let's Encrypt
• readonly mode with a query execution timeout
• Custom GRASS style
• Custom guides

Public Hetionet Neo4j Instance

MATCH path =
  // Specify the type of path to match
  (n0:Disease)-[e1:ASSOCIATES_DaG]-(n1:Gene)-[:INTERACTS_GiG]-
  (n2:Gene)-[:PARTICIPATES_GpBP]-(n3:BiologicalProcess)
WHERE
  // Specify the source and target nodes
  n0.name = 'multiple sclerosis' AND
  n3.name = 'retina layer formation'
  // Require GWAS support for the
  // Disease-associates-Gene relationship
  AND 'GWAS Catalog' in e1.sources
  // Require the interacting gene to be
  // upregulated in a relevant tissue
  AND exists(
    (n0)-[:LOCALIZES_DlA]-(:Anatomy)-[:UPREGULATES_AuG]-(n2))
RETURN path

How could multiple sclerosis could affect retina layer formation?

More queries at thinklab.com/d/220

Project Rephetio contributions on Thinklab

(see thinklab.com/p/rephetio/leaderboard)

Project Rephetio: drug repurposing predictions

• Hetionet v1.0 contains:
  • 1,538 connected compounds
  • 136 connected diseases
  • 209,168 compound–disease pairs
  • 755 treatments
• Systematic drug repurposing:
  • Compare the therapeutic utility of data types
  • Identify the mechanisms of drug efficacy
  • Predict the probability of treatment for all 209,168 compound–disease pairs (het.io/repurpose)

Systematic integration of biomedical knowledge prioritizes drugs for repurposing
Daniel S Himmelstein, Antoine Lizee, Christine Hessler, Leo Brueggeman, Sabrina L Chen, Dexter Hadley, Ari Green, Pouya Khankhanian, Sergio E Baranzini
eLife (2017) https://doi.org/10.7554/eLife.26726

Machine learning methodology

Predictions succeed at prioritizing known treatments

• disease modifying treatments
  +755, −208,413
  AUROC = 97.4%

Predictions succeed at prioritizing experimental treatments

• disease modifying treatments
  +755, −208,413
  AUROC = 97.4%

• treatments in clinical trials
  +5,594, −202,186
  AUROC = 70.0%

1,206 compound–disease metapaths (length ≤ 4)

1. Upper tier:
   traditional pharmacology
2. Upper-middle tier:
   traditionally biomedicine, but newer in drug efficacy
3. Lower-middle tier:
   genome-wide / high-throughput data sources
4. Lower tier:
   cellular components

Browse at het.io/repurpose/metapaths.html

Project Rephetio: Does bupropion treat nicotine dependence?

• Bupropion was first approved for depression in 1985
   
• In 1997, bupropion was approved for smoking cessation
   
• Can we predict this repurposing from Hetionet? The prediction was:
  • 99.5th percentile for nicotine dependence
  • probability 2.50-fold greater than null

Compound–causes–SideEffect–causes–Compound–treats–Disease

Compound–binds–Gene–binds–Compound–treats–Disease

Compound–binds–Gene–associates–Disease

Compound–binds–Gene–participates–Pathway–participates–Disease

MATCH path = (n0:Compound)-[:BINDS_CbG]-(n1)-[:PARTICIPATES_GpPW]-
  (n2)-[:PARTICIPATES_GpPW]-(n3)-[:ASSOCIATES_DaG]-(n4:Disease)
USING JOIN ON n2
WHERE n0.name = 'Bupropion'
  AND n4.name = 'nicotine dependence'
  AND n1 <> n3
WITH
[
  size((n0)-[:BINDS_CbG]-()),
  size(()-[:BINDS_CbG]-(n1)),
  size((n1)-[:PARTICIPATES_GpPW]-()),
  size(()-[:PARTICIPATES_GpPW]-(n2)),
  size((n2)-[:PARTICIPATES_GpPW]-()),
  size(()-[:PARTICIPATES_GpPW]-(n3)),
  size((n3)-[:ASSOCIATES_DaG]-()),
  size(()-[:ASSOCIATES_DaG]-(n4))
] AS degrees, path
RETURN
  path,
  reduce(pdp = 1.0, d in degrees| pdp * d ^ -0.4) AS path_weight
ORDER BY path_weight DESC
LIMIT 10

Cypher query to find the top CbGbPWaD paths

Browse all predictions at het.io/repurpose. Discuss at thinklab.com/d/224

Top 100 epilepsy predictions & their chemical structure

Discuss at thinklab.com/d/224#5

Top 100 epilepsy predictions & their drug targets

Discuss at thinklab.com/d/230#14

• Hetionet (≤ v1.0) integrated data from 31 resources:
  • 5 United States Government works
  • 12 openly licensed
  • 4 non-commercial use only
  • 9 were all rights reserved
  • 1 explicitly & contractually forbid reuse
• Requested permission for 11 resources:
  • median time to first response was 16 days
  • 2 affirmative responses
• Other considerations:
  • who owns data
  • incompatibilities: share alike vs non-commercial
  • copyright status of data & fair use
• Solution: license attribute per node/relationship

Legal barriers to data reuse

Recommendations:

1. release data under an open license
2. University researchers: commit to open in your resource sharing plan

hetmech

https://github.com/greenelab/hetmech

the hetnet search engine

https://zietzm.github.io/Vagelos2017/

Future: all biomedical knowledge in a single network

https://github.com/greenelab/snorkeling

• Teach computers how to read the literature and extract knowledge.
   
• Continuously and automatically refine and grow the hetnet.
   
• Free from any legal restrictions on reuse.