Integrate all: hetnets in human disease.

September 24, 2015

Smilow 10-120, UPenn

dhimmel on:

Greene Lab Interview

—Daniel Himmelstein

Moore Lab Lunch

Jesse's Tavern

August, 2007

Baranzini Lab @ UCSF

Sandler Neurosciences Center

Sergio

PhD in Biological & Medical Informatics

Hodgkin's lymphoma is genetically closer to autoimmune diseases than solid cancers

Founding Insight

context
bioinformatics → data explosion
goal
mine the data to advance human health
problem
high-throughput data tends to predict weakly
remedy
combine diverse datasets into a strong predictor
method
heterogeneous network (hetnet) edge prediction

predicting disease-associated genes

wealth of GWAS associations

integrate diverse data to provide context

network of pathogenesis:

18 metanodes
40,343 nodes
19 metaedges
1,608,168 edges

type is essential when operating on hetnets

metapath-based approach

feature extraction: the DWPC

learning to classify:

regularized logistic regression

performance and permuation

mechanisms of pathogenesis:

comparing gene set collections

mechanisms of pathogenesis:

comparing metapaths

Predicting withheld MS associations

Novel MS associations

Extra Slides Below

gene extraction from the GWAS Catalog

Feature redundancy

Metanodes (node types)

Metaedges (edge types)

robustness

Network subsampling

drug repurposing

massively collaborative open science

22 reviewers, 47 discussions, 266 comments

Mining knowledge from 69 years of biomedical publication

MEDLINE: curators annotate paper topics
21 million articles
since 1946
5,594 journals
cooccurence of two topics indicates a relation
disease–symptom cooccurence in 363,928 articles, 696,252 for disease–anatomy

Discuss on

anatomy

symptom

Mining MEDLINE for disease context

LINCS L1000: ~20,000 small molecules

We combine all signatures for each DrugBank compound to get a consensus signature

Brueggeman

transcriptional signatures discriminate

diuretics (DR) & anti-Inflammatories (AA)

← genes →

Disease-specific models uncover therapeutic signatures

MEDI-HPS:
- RxNorm
- MedlinePlus
- SIDER 2
- Wikipedia
ehrlink: linked data from health records
LabeledIn: expert and MTurk curated drug labels
PREDICT:
- UMLS links
- drugs.com
- drug labels

catalog of indications

aggregated 4 databases:

yielding 1,388 indications

drug	disease	ajg	csh	eq
Acetylsalicylic acid	systemic lupus erythematosus	DM	DM	1
Alprazolam	systemic scleroderma	SYM	SYM	1
Baclofen	multiple sclerosis	SYM	SYM	1
Bupropion	panic disorder	SYM	DM	0
Captopril	rheumatoid arthritis	NOT	NOT	1
Cisplatin	hematologic cancer	DM	DM	1
Cladribine	hematologic cancer	DM	DM	1
Clopidogrel	coronary artery disease	DM	DM	1
Cocaine	dental caries	NOT	SYM	0

class	ajh	csh
DM	26	32
SYM	20	17
NOT	4	1

curation pilot results

~58% disease modifying
~37% symptomatic
~5% non-indications
discuss on

66% ✓

legal forays of a modern biodata scientist

Network contains data from 28 resources:

12 lack any licensing information
10 use standard licenses
6 use custom licenses
3 resources are publication supplements
6 forbid commercial use
2 forbid any redistribution of the data

Project is open notebook and maximally reusable

After a 5000+ word discussion:

identified a path to compliance that minimizes sacrifice
suggestion: release your data under CC0

results

metapath	nonzero	auroc
CcSEcCiD	0.590	0.897
CiDiCiD	0.255	0.840
CiDaGaD	0.411	0.820
CtGtCiD	0.227	0.807
CiDpSpD	0.410	0.801
CiDlAlD	0.400	0.797
CtGiGaD	0.335	0.710
CuG<kuGaD	0.425	0.692
CtGvD	0.009	0.514

Results for all metapaths ≤ length 3

feature performance

(including symptomatic indications)

model

lasso auroc = 0.967

collinearity

subset of all 261 features

hetnets

data integration
- scalable
- intuitive
edge prediction
- disease-associated genes
- drug indications
powerful
- upfront integration cost
- subsequent convenience

>>> import phd