General Session IVUtilizing Big Data to Better Understand

Environment-related Research:Analytics and Genomic Health Disparities

JOIN OUR SOCIAL MEDIA DISCUSSIONS

#XUDisparitiesCollabs

Accreditation

Participation in this activity earns 1.50 contact hours. To receive credit, participants must complete an evaluation form at the

end of this session.

UAN: 0024-0000-17-006-L04-P

DISCLOSURE STATEMENTSpeakers for this session have nothing to disclose. Any updates

in disclosure will be made from the podium.

Michael A. Langston, PhDSPEAKER

Health Disparities Research: The Role of Big Data and Scalable Combinatorial Methods in

Latent Relationship Discovery

Xavier University of Louisiana College of PharmacyTenth Health Disparities Conference

Michael A. LangstonProfessor

Electrical Engineering and Computer ScienceUniversity of Tennessee

17 March 2017

55

Historical Perspective

Some have likened the male Fiddler Crab tothe current state of research balance.

Considerable effort has goneinto powerful omics tools.

Relatively little focus has beenshed on the analysis of riskfactors for health disparities.

Christopher P. Wild, Director, Int’lAgency for Research on Cancer

56

Some Things We Have Learned From Omics

The situation is seldom as simple as it seems at first• DNA is not the whole story• tissue types, environments, …

The central dogma is just a starting point• one gene may yield many proteins• then there is post translational modification• and also methylation, chromatin remodeling, …

One needs a “systems” approach• How do biological entities function at all levels of scale?• Linkage, communication, granularity and networks (graphs!)

Even crude tools can be useful• Knock outs, knock ins, knock downs, …• As reverse engineering methods, are they really so sophisticated?

57

A Confluence of Resources for Health Disparities

Environmental Health Disparities Data

Deep Domain Knowledge

Novel Combinatorial Algorithms

High Performance Platforms

All Focused on a Central Problem

Understanding Health Disparities (e.g., in the

Context of the Exposome)

58

General Approach

Resources unique to the University of Tennessee• Hub for deep work in graph-theoretical algorithms• Home to well-known and widely used software products, such as MPI• Next door to ORNL, with access to leadership-class machines like Titan

Main Research Goal• Uncover relationships deeply hidden within high dimensional data• Exploit cutting-edge graph-theoretical algorithms, coupled with high

performance computing and extensive domain knowledge• Extract signal from noise in so-called real-world big data

Focus on Optimization• Solve computationally recalcitrant (usually NP-hard) problems• Employ mathematical abstraction, computational breakthroughs and extreme

scalability• Distill and extract dense subgraphs and patterns of connectivity• Augment with statistical tools whenever possible

59

General Approach

Data• Expensive, difficult and/or time-consuming to obtain

Problems• Infeasible to solve exactly without advanced algorithms/software/hardware

Keys to Scalable Analytics• State-of-the-art graph-theoretical algorithms• Fixed Parameter Tractability

o Pioneering work of Fellows-Langston going back over 25 yearso Now a cottage industry around the world

• High Performance Computing• Duality and polynomial-time transforms (e.g., to/from SAT and ILP)• Statistical tools (e.g., Bayesian, Bonferroni and FDR) for pre- and post-

processing• Domain knowledge – know your data and use its structure to your advantage

60

General Approach

Input to Sponsor

Heterogeneous Data

Relevance Networks

Dense Subgraphs

Testable Hypotheses

Multi-Level Modeling, Dimension Reduction, Correlation, Thresholding, Contrasts,

Spectral Methods

Graph-Theoretical Algorithms, Clique-Centric Tools, High Performance Computing, Clustering, Anchoring

Differential Topology, Case/Control Analysis, Near-Neighborhood Expansion, Structural

Equation Modeling

Bayesian Refinement, Validation and Testing, User Participation, On-Site Study,

Feedback for Continued Analysis

A Sample Toolchain (feedback loops not shown)

Supercomputers Often Needed• for example Edison, an NERSC Cray XC30 system• 133,824 compute cores, 2.57 petaflops, 357 TB of memory• Also machines at DOD, ORNL and XSEDE (formerly the TeraGrid)

61

Practicalities and Implementations

0

200

400

600

800

1000

1200

1400

0 200 400 600 800 1000 1200

spee

dup

# of processors

optimum (linear) speedup

dynamic load balancing (estra-30)

dynamic load balancing (folic-30)

dynamic load balancing (avg)

Workload Balancing

62

Practicalities and Implementations

63

Health Disparities and the Sampling Question

A Sampling Spectrum – Where to Begin?

Whole Blood PBMC CD4 Th2

Purity of SamplingBut Information Overload?

PrecisionBut Loss of Bio Function?

The ideal question may be: where are the best biomarkers?But the real question is often: what is available?

Meta-Application: Data Clustering

Average Quartile

Small (3-10 genes) Medium (11-100 genes) Large (101-1000 genes)

Clustering Method QuartileBAT5

Jaccard Quartile BAT5 Jaccard Quartile BAT5 Jaccard

K-Clique Communities 1.00 1 0.7531 1 0.4465 1 0.4915

Maximal Clique 1.00 1 0.8433 1 0.4081 0.0000

Paraclique 1.00 1 0.7576 1 0.4285 1 0.4169

Ward (H) 1.33 2 0.5782 1 0.4011 1 0.5723

CAST 1.67 1 0.7455 3 0.3146 1 0.4994

QT Clust 2.00 2 0.5473 2 0.3670 2 0.3944

Complete (H) 2.33 3 0.3933 2 0.3677 2 0.3419

NNN 2.67 2 0.5521 2 0.3705 4 0.2406

K-Means 3.00 4 0.2573 3 0.3015 2 0.3463

SOM 3.00 4 0.3260 2 0.3286 3 0.3282

WGCNA 3.00 3 0.4391 3 0.3106 3 0.2949

Average (H) 3.33 3 0.4087 4 0.2792 3 0.3037

McQuitty (H) 3.33 3 0.4594 3 0.3065 4 0.2868

SAMBA 3.50 0.0000 4 0.1860 3 0.3298

CLICK 4.00 4 0.0339 4 0.1453 4 0.2817

Clustering Saccharomyces cerevisiae data,algorithms ranked by quartile comparisons

64

Biological Application: Allergic Rhinitis

Transcriptomic Data Description• Mikael Benson, Göteborg, Sweden, 56 patients and 39 controls• Affymetrix HU133 arrays• roughly 33,000 genes (~1B interactions)• nasal secretions, lymphocytes, skin• hay fever, eczema

Preprocessing• MAS5.0• log transformed• replicates averaged• centered around zero with z scores• probesets with consistently low expression levels removed

Threshold Selection• spectral graph theory used to set and balance graph densities• AFFX spots retained for quality control

0

500000

1000000

1500000

2000000

2500000

Correlation Value

Freq

uenc

y

PatientControl

65

Clique profiles using the five most highly represented genes:

Control Patient

Gene Symbol Clique membership Gene Symbol Clique membership

UBE1C 29% FGFR2 66%

RANBP6 27% NFIB 65%

DKFZP564O123 26% PPL 64%

SLC25A13 24% FGFR3 64%

GTPBP4 21% CDH3 56%

ribosomal or RNA-related T-lymphocytes or epithelial cells

Applied differential screens, then chromatin immunoprecipitation.Sample Result: Discovered a novel and key role for ITK (IL2-inducible T-cell kinase).

66

Biological Application: Allergic Rhinitis

0

5

10

15

20

25

Patient-Control

Patient-Patient

Control-Control

A real challenge: discordant monozygotic twins• most types of omics data fail• so use epigenetics - methylation biomarkers• captures environmental exposures & randomization• best separation we’ve seen

67

Biological Application: Allergic Rhinitis

Geo-Resource Application:North Sea Historical Data

Diverse sets of oceanographic data since the 1950sSpatial analysis via paraclique supports the North Sea flushing rate

68

Hurricane Katrina’s impact on minority populations, as measured bymortality, morbidity and displacement

69

Health Disparities Application:Hurricane Katrina

Sample Gulf Coast Analysis: Mortality Quadrants circa 2005

70

Health Disparities Application:Hurricane Katrina

Meat Markets

Total Arrests

Emergency Relief Services

EducationalServices

Total Birth

Convenience Stores

MentalHealth

Centers

Nursing Centers

Non Mental Health MDs

Prenatal Care

Residential Retardation

Centers

Dentists

Super-markets

Fish Markets

PoliticalAssociations

Grant MakingAssociations

LaborAssociations

# of Business Establishments

Massive/Severe Damage Major Damage

# of Housing Units

Paraclique of21 correlatesin Gulf Coastcounties and

parishes

71

Health Disparities Application:Hurricane Katrina

72

• preterm defined as birth at less than 37 weeks• significant risk factor for infant death• paracliques, latent factors, Bayesian analysis

Health Disparities Application:Prematurity

73

Data• 3143 county-equivalents in a four-case study design: BM, WM, BF, WF

Cigarette consumption a contributing factor for all four groupsBut smoking alone

• does not explain differences in race x gender mortality disparities• its relative contribution to lung cancer mortality and mortality disparities

• relatively high for WM and WF• much less so for BM and BF

Need to uncover pathways through which environmental exposures• increase risk for lung cancer mortality• lead to race x gender disparities• help identify populations at greatest risk• targeting appropriate interventions

Health Disparities Application:Lung Cancer

Recent Software Releases

74

EntropyExplorer (R package)• Efficient and convenient computation and comparison of differential

Shannon entropy, differential coefficient of variation and differential expression

GrAPPA (website)• Plots, histograms, intensity diagrams and various normalizations• Correlation computations (Pearson, Spearman, Mutual Information,

Jaccard, others)• Graph creation and decomposition• Numerous clustering algorithms• Visualization tools

GeneWeaver (website)• Classify behavioral traits via shared biomolecular substrates, and

validate classifications through tests of pleiotropic effects

Health Disparities Brings with It Serious Data Issues• NDAs• CDC• NIH & HIPAA• DOD &TCPs• IRB• Non-de-identified data … even so-called de-identified data!

Support Gratefully Acknowledged• Domestic: AFRL, DOE (BER), NIH (NIAAA, NIDA, NIMHD), NRO• International: ARC, EU

Thanks for Listening!

Closing Remarks

75