Estimating AutoAntibody Signatures to Detect Autoimmune Disease - - PowerPoint PPT Presentation

▶
estimating autoantibody signatures to detect autoimmune
SMART_READER_LITE
LIVE PREVIEW

Estimating AutoAntibody Signatures to Detect Autoimmune Disease - - PowerPoint PPT Presentation

Jul 08, 2023 107 likes •917 views

Individualized Health Background Method Application Summary Estimating AutoAntibody Signatures to Detect Autoimmune Disease Patient Subsets Zhenke Wu Assistant Professor Department of Biostatistics, University of Michigan 27 July 2017 The

slide-1
SLIDE 1

Individualized Health Background Method Application Summary

Estimating AutoAntibody Signatures to Detect Autoimmune Disease Patient Subsets

Zhenke Wu

Assistant Professor Department of Biostatistics, University of Michigan

27 July 2017 The 62nd Annual International Biometric Society Meeting of the Brazilian Region (RBras 2017) R package: spotgear https://github.com/zhenkewu/spotgear

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 1 / 30

slide-2
SLIDE 2

Individualized Health Background Method Application Summary

Common Questions on Individual and Population Health

1.

  • a. What is the person’s health state

given health measurements?

  • b. What is the population distribution
  • f health states?

(Wu et al., 2015, JRSS-C; Wu and Zeger, 2016a,b)

2.

  • a. What is the person’s health

trajectory?

  • b. What is the population’s

characteristics of health trajectory?

  • 3. Does a particular intervention improve

health - on average/for a particular person? (Wu et al., 2014, Biometrics;

Frangakis, Qian, Wu, Diaz, 2015, Biometrics)

  • 4. Are interventions being used
  • ptimally?

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 2 / 30

slide-3
SLIDE 3

Individualized Health Background Method Application Summary

Example I

Pneumonia Etiology Research for Child Health (PERCH)

Background:

  • > 30 possible infectious causes
  • Difficult to directly observe

Goal:

  • Population disease etiology estimation
  • Individual diagnosis

Study details:

  • $40-mil, Gates-funded 7-country study;

Sites at Sub-Saharan Africa and South Asia

  • Diverse measures; variable precisions
  • ∼5,000 cases and ∼5,000 controls

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 3 / 30

slide-4
SLIDE 4

Individualized Health Background Method Application Summary

Measurements of Different Quality

cases

(~5,000)

controls

(~5,000)

Lung Infection

NA

  • *NP: nasopharyngeal; PCR: polymerase chain reaction; LA: lung aspirate

NA NA

𝐽𝑀

𝑗

Nasopharyngeal PCR

Blood Culture Lung aspirate 𝑁𝑗

𝑇

Bronze- Standard (BrS) Silver- Standard (SS) Gold- Standard (GS)

(𝜄, 𝜔)

Latent Health State Measurements Measurement Precisions Specimen (S)

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 4 / 30

slide-5
SLIDE 5

Individualized Health Background Method Application Summary Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 5 / 30

slide-6
SLIDE 6

Individualized Health Background Method Application Summary

Example II: Raw Data

Gel Electrophoresis Autoradiography; 20 Samples

Raw Image

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 6 / 30

slide-7
SLIDE 7

Individualized Health Background Method Application Summary

Example II: Raw Data

Gel Electrophoresis Autoradiography; 20 Samples

Raw Image

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 6 / 30

slide-8
SLIDE 8

Individualized Health Background Method Application Summary

Example II: Raw Data

Gel Electrophoresis Autoradiography; 20 Samples

Raw Image

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 6 / 30

slide-9
SLIDE 9

Individualized Health Background Method Application Summary

Example II: Raw Data

Gel Electrophoresis Autoradiography; 20 Samples

Raw Image

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 6 / 30

slide-10
SLIDE 10

Individualized Health Background Method Application Summary

Example II: Raw Data

Gel Electrophoresis Autoradiography; 20 Samples

Raw Image

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 6 / 30

slide-11
SLIDE 11

Individualized Health Background Method Application Summary

Example II: Raw Data

Gel Electrophoresis Autoradiography; 20 Samples

Raw Image Hand-picked Bands “|”

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 6 / 30

slide-12
SLIDE 12

Individualized Health Background Method Application Summary

Roadmap

Laboratory Data (AutoAntibody Profile)

Enroll Autoimmune Disease Patients Obtain Serum Samples (contain AutoAntibodies) Add cell lysate (contain proteins that AutoAntibodies bind, e.g., Ro60, RNP , PL-7 and centromere) Visualize the IP bands (Autoradiography) Obtain a mixture of binded AutoAntibodies, i.e. Immuno- Precipitation (IP) Separate AutoAntibodies by Gel Electrophoresis (SDS-PAGE gels)

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 7 / 30

slide-13
SLIDE 13

Individualized Health Background Method Application Summary

Roadmap

Laboratory Data (AutoAntibody Profile)

Enroll Autoimmune Disease Patients Obtain Serum Samples (contain AutoAntibodies) Add cell lysate (contain proteins that AutoAntibodies bind, e.g., Ro60, RNP , PL-7 and centromere) Visualize the IP bands (Autoradiography) Obtain a mixture of binded AutoAntibodies, i.e. Immuno- Precipitation (IP) Separate AutoAntibodies by Gel Electrophoresis (SDS-PAGE gels)

Data Science (Patient Subsetting)

Enroll Autoimmune Disease Patients Obtain Serum Samples (contain AutoAntibodies)

Patient Subsetting with distinct AutoAntibody Signatures

Add cell lysate (contain proteins that AutoAntibodies bind, e.g., Ro60, RNP , PL-7 and centromere)

Visualize the IP bands (Autoradiography) Obtain a mixture of binded AutoAntibodies, i.e. Immuno- Precipitation (IP) Separate AutoAntibodies by Gel Electrophoresis (SDS-PAGE gels)

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 7 / 30

slide-14
SLIDE 14

Individualized Health Background Method Application Summary

Roadmap

Laboratory Data (AutoAntibody Profile)

Enroll Autoimmune Disease Patients Obtain Serum Samples (contain AutoAntibodies) Add cell lysate (contain proteins that AutoAntibodies bind, e.g., Ro60, RNP , PL-7 and centromere) Visualize the IP bands (Autoradiography) Obtain a mixture of binded AutoAntibodies, i.e. Immuno- Precipitation (IP) Separate AutoAntibodies by Gel Electrophoresis (SDS-PAGE gels)

Data Science (Patient Subsetting)

Enroll Autoimmune Disease Patients Obtain Serum Samples (contain AutoAntibodies)

Patient Subsetting with distinct AutoAntibody Signatures

Add cell lysate (contain proteins that AutoAntibodies bind, e.g., Ro60, RNP , PL-7 and centromere)

Visualize the IP bands (Autoradiography) Obtain a mixture of binded AutoAntibodies, i.e. Immuno- Precipitation (IP) Separate AutoAntibodies by Gel Electrophoresis (SDS-PAGE gels)

Phenotypic Data

1) Johns Hopkins Autoimmune Disease Patient Cohort with phenotype data (cancer type, organ functions, etc.) 2) Population Cancer Registry Data as reference (Surveillance, Epidemiology and End Results - SEER)

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 7 / 30

slide-15
SLIDE 15

Individualized Health Background Method Application Summary

Roadmap

Laboratory Data (AutoAntibody Profile)

Enroll Autoimmune Disease Patients Obtain Serum Samples (contain AutoAntibodies) Add cell lysate (contain proteins that AutoAntibodies bind, e.g., Ro60, RNP , PL-7 and centromere) Visualize the IP bands (Autoradiography) Obtain a mixture of binded AutoAntibodies, i.e. Immuno- Precipitation (IP) Separate AutoAntibodies by Gel Electrophoresis (SDS-PAGE gels)

Data Science (Patient Subsetting)

Enroll Autoimmune Disease Patients Obtain Serum Samples (contain AutoAntibodies)

Patient Subsetting with distinct AutoAntibody Signatures

Add cell lysate (contain proteins that AutoAntibodies bind, e.g., Ro60, RNP , PL-7 and centromere)

Visualize the IP bands (Autoradiography) Obtain a mixture of binded AutoAntibodies, i.e. Immuno- Precipitation (IP) Separate AutoAntibodies by Gel Electrophoresis (SDS-PAGE gels)

Phenotypic Data

1) Johns Hopkins Autoimmune Disease Patient Cohort with phenotype data (cancer type, organ functions, etc.) 2) Population Cancer Registry Data as reference (Surveillance, Epidemiology and End Results - SEER) Risk stratification (with cancer phenotypes) Predict the magnitude & timing

  • f cancer risks in each subset,

e.g., “Which AutoAntibody signatures protect an individual against cancer?“

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 7 / 30

slide-16
SLIDE 16

Individualized Health Background Method Application Summary

Roadmap

Laboratory Data (AutoAntibody Profile)

Enroll Autoimmune Disease Patients Obtain Serum Samples (contain AutoAntibodies) Add cell lysate (contain proteins that AutoAntibodies bind, e.g., Ro60, RNP , PL-7 and centromere) Visualize the IP bands (Autoradiography) Obtain a mixture of binded AutoAntibodies, i.e. Immuno- Precipitation (IP) Separate AutoAntibodies by Gel Electrophoresis (SDS-PAGE gels)

Data Science (Patient Subsetting)

Enroll Autoimmune Disease Patients Obtain Serum Samples (contain AutoAntibodies)

Patient Subsetting with distinct AutoAntibody Signatures

Add cell lysate (contain proteins that AutoAntibodies bind, e.g., Ro60, RNP , PL-7 and centromere)

Visualize the IP bands (Autoradiography) Obtain a mixture of binded AutoAntibodies, i.e. Immuno- Precipitation (IP) Separate AutoAntibodies by Gel Electrophoresis (SDS-PAGE gels)

Phenotypic Data

1) Johns Hopkins Autoimmune Disease Patient Cohort with phenotype data (cancer type, organ functions, etc.) 2) Population Cancer Registry Data as reference (Surveillance, Epidemiology and End Results - SEER) Risk stratification (with cancer phenotypes) Predict the magnitude & timing

  • f cancer risks in each subset,

e.g., “Which AutoAntibody signatures protect an individual against cancer?“

This Talk

Hierarchical Bayesian Models for Preprocessing

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 7 / 30

slide-17
SLIDE 17

Individualized Health Background Method Application Summary

Autoimmune Diseases and AutoAntibody Signatures

  • Etiology of Autoimmue Diseases:

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 8 / 30

slide-18
SLIDE 18

Individualized Health Background Method Application Summary

Autoimmune Diseases and AutoAntibody Signatures

  • Etiology of Autoimmue Diseases:
  • Human immune system’s responses to autoantigens;

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 8 / 30

slide-19
SLIDE 19

Individualized Health Background Method Application Summary

Autoimmune Diseases and AutoAntibody Signatures

  • Etiology of Autoimmue Diseases:
  • Human immune system’s responses to autoantigens;
  • The body produces specific autoantibodies that target these

autoantigens but also cause tissue damage

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 8 / 30

slide-20
SLIDE 20

Individualized Health Background Method Application Summary

Autoimmune Diseases and AutoAntibody Signatures

  • Etiology of Autoimmue Diseases:
  • Human immune system’s responses to autoantigens;
  • The body produces specific autoantibodies that target these

autoantigens but also cause tissue damage

  • Heterogeneity: The autoantibody composition is strikingly

different among patients

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 8 / 30

slide-21
SLIDE 21

Individualized Health Background Method Application Summary

Autoimmune Diseases and AutoAntibody Signatures

  • Etiology of Autoimmue Diseases:
  • Human immune system’s responses to autoantigens;
  • The body produces specific autoantibodies that target these

autoantigens but also cause tissue damage

  • Heterogeneity: The autoantibody composition is strikingly

different among patients

  • Long-term clinical objective: find autoantibody signature that

subsets autoimmune disease patients into groups with more homogeneous phenotypes and trajectories

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 8 / 30

slide-22
SLIDE 22

Individualized Health Background Method Application Summary

Autoimmune Diseases and AutoAntibody Signatures

  • Etiology of Autoimmue Diseases:
  • Human immune system’s responses to autoantigens;
  • The body produces specific autoantibodies that target these

autoantigens but also cause tissue damage

  • Heterogeneity: The autoantibody composition is strikingly

different among patients

  • Long-term clinical objective: find autoantibody signature that

subsets autoimmune disease patients into groups with more homogeneous phenotypes and trajectories

  • Measurements:

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 8 / 30

slide-23
SLIDE 23

Individualized Health Background Method Application Summary

Autoimmune Diseases and AutoAntibody Signatures

  • Etiology of Autoimmue Diseases:
  • Human immune system’s responses to autoantigens;
  • The body produces specific autoantibodies that target these

autoantigens but also cause tissue damage

  • Heterogeneity: The autoantibody composition is strikingly

different among patients

  • Long-term clinical objective: find autoantibody signature that

subsets autoimmune disease patients into groups with more homogeneous phenotypes and trajectories

  • Measurements: Gel Electrophoresis Autoradiography (GEA)

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 8 / 30

slide-24
SLIDE 24

Individualized Health Background Method Application Summary

Gel Electrophoresis Autoradiography (GEA)

A technique to visualize the abundance of molecules or fragments

  • f molecules that have been radioactively labeled.
  • Can generate 100s of possibilities of band patterns

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 9 / 30

slide-25
SLIDE 25

Individualized Health Background Method Application Summary

Gel Electrophoresis Autoradiography (GEA)

A technique to visualize the abundance of molecules or fragments

  • f molecules that have been radioactively labeled.
  • Can generate 100s of possibilities of band patterns
  • Can be tested and validated using commercially available line

immunoblot assay (EuroImmun; Systemic Sclerosis (Nucleoli) profile)

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 9 / 30

slide-26
SLIDE 26

Individualized Health Background Method Application Summary

Gel Electrophoresis Autoradiography (GEA)

A technique to visualize the abundance of molecules or fragments

  • f molecules that have been radioactively labeled.
  • Can generate 100s of possibilities of band patterns
  • Can be tested and validated using commercially available line

immunoblot assay (EuroImmun; Systemic Sclerosis (Nucleoli) profile)

  • Gap: Onerous and expensive to validate; Need a method to

greatly simplify autoantibody profile discovery

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 9 / 30

slide-27
SLIDE 27

Individualized Health Background Method Application Summary

Gel Electrophoresis Autoradiography (GEA)

A technique to visualize the abundance of molecules or fragments

  • f molecules that have been radioactively labeled.
  • Can generate 100s of possibilities of band patterns
  • Can be tested and validated using commercially available line

immunoblot assay (EuroImmun; Systemic Sclerosis (Nucleoli) profile)

  • Gap: Onerous and expensive to validate; Need a method to

greatly simplify autoantibody profile discovery

  • Solution: Pre-filtering to define subgroups with similar

specificities based on the bands observed by GEA

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 9 / 30

slide-28
SLIDE 28

Individualized Health Background Method Application Summary

Automated Pipeline for Autoimmune Disease Subsetting

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 10 / 30

slide-29
SLIDE 29

Individualized Health Background Method Application Summary

Step I-A: Automated Peak Detection

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 11 / 30

slide-30
SLIDE 30

Individualized Health Background Method Application Summary

Step I-A: Automated Peak Detection

Overlaid against gel image; “∗” for detected peaks

  • ugi: lane number for lane i = 1, . . . , Ng, gel g = 1, . . . , G

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 12 / 30

slide-31
SLIDE 31

Individualized Health Background Method Application Summary

Step I-A: Automated Peak Detection

Overlaid against gel image; “∗” for detected peaks

  • ugi: lane number for lane i = 1, . . . , Ng, gel g = 1, . . . , G
  • Tgij: location for the j-th peak (“∗”), j = 1, . . . , Jgi, for lane

i, gel g

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 12 / 30

slide-32
SLIDE 32

Individualized Health Background Method Application Summary

Step I-B: Batch Effect Correction

Must address before meaningful subgrouping

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 13 / 30

slide-33
SLIDE 33

Individualized Health Background Method Application Summary

Warping Examples

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 14 / 30

slide-34
SLIDE 34

Individualized Health Background Method Application Summary

Step I-B: Batch Effect Correction

Piecewise Linear Warping by Reference Marker Molecules

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 15 / 30

slide-35
SLIDE 35

Individualized Health Background Method Application Summary

Step I-C: Align the peaks

”Which “+” do the peaks “•” belong?”

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 16 / 30

slide-36
SLIDE 36

Individualized Health Background Method Application Summary

Step I-C: Two-Dimensional De-Warping

  • The physical process of autoradiography could cause image

deformation

  • Challenges
  • In general, few light-weight proteins on the right side of the image; If

we don’t see bands, how to align? Solution: align to a grid of protein landmarks and assume smoothness

  • f warping

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 17 / 30

slide-37
SLIDE 37

Individualized Health Background Method Application Summary

Step I-C: Two-Dimensional De-Warping

  • The physical process of autoradiography could cause image

deformation

  • Challenges
  • In general, few light-weight proteins on the right side of the image; If

we don’t see bands, how to align? Solution: align to a grid of protein landmarks and assume smoothness

  • f warping
  • Ubiquitous proteins (e.g., actin) on multiple gels must be aligned.

Solution: Discretized non-homogeneous Poisson process with shared intensity across gels

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 17 / 30

slide-38
SLIDE 38

Individualized Health Background Method Application Summary

Step I-C: Two-Dimensional De-Warping

  • The physical process of autoradiography could cause image

deformation

  • Challenges
  • In general, few light-weight proteins on the right side of the image; If

we don’t see bands, how to align? Solution: align to a grid of protein landmarks and assume smoothness

  • f warping
  • Ubiquitous proteins (e.g., actin) on multiple gels must be aligned.

Solution: Discretized non-homogeneous Poisson process with shared intensity across gels

  • The observed peak locations are noisy.

Solution: Gaussian noise around the true location

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 17 / 30

slide-39
SLIDE 39

Individualized Health Background Method Application Summary

Step I-C: Model for 2-Dimensional Image Dewarping

Prior on the peak-to-landmark indicators

  • Peak-to-landmark Indicators:

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 18 / 30

slide-40
SLIDE 40

Individualized Health Background Method Application Summary

Step I-C: Model for 2-Dimensional Image Dewarping

Prior on the peak-to-landmark indicators

  • Peak-to-landmark Indicators:
  • 1. Zgij ∈ {1, . . . , L}, j = 1, . . . , Jgi (match a “∗” to a “+”), e.g.,

Zgij = 3 means the peak is matched to Landmark 3

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 18 / 30

slide-41
SLIDE 41

Individualized Health Background Method Application Summary

Step I-C: Model for 2-Dimensional Image Dewarping

Prior on the peak-to-landmark indicators

  • Peak-to-landmark Indicators:
  • 1. Zgij ∈ {1, . . . , L}, j = 1, . . . , Jgi (match a “∗” to a “+”), e.g.,

Zgij = 3 means the peak is matched to Landmark 3

  • 2. Constrain Zgi,j−1 ≤ Zgij to prevent reverse matching

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 18 / 30

slide-42
SLIDE 42

Individualized Health Background Method Application Summary

Step I-C: Model for 2-Dimensional Image Dewarping

Prior on the peak-to-landmark indicators

  • Peak-to-landmark Indicators:
  • 1. Zgij ∈ {1, . . . , L}, j = 1, . . . , Jgi (match a “∗” to a “+”), e.g.,

Zgij = 3 means the peak is matched to Landmark 3

  • 2. Constrain Zgi,j−1 ≤ Zgij to prevent reverse matching
  • Bayesian Model for Aligning Peaks to Landmarks

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 18 / 30

slide-43
SLIDE 43

Individualized Health Background Method Application Summary

Step I-C: Model for 2-Dimensional Image Dewarping

Prior on the peak-to-landmark indicators

  • Peak-to-landmark Indicators:
  • 1. Zgij ∈ {1, . . . , L}, j = 1, . . . , Jgi (match a “∗” to a “+”), e.g.,

Zgij = 3 means the peak is matched to Landmark 3

  • 2. Constrain Zgi,j−1 ≤ Zgij to prevent reverse matching
  • Bayesian Model for Aligning Peaks to Landmarks
  • Number of observed peaks in lane i, gel g:

Jgi

d

∼ Poisson(Λ)

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 18 / 30

slide-44
SLIDE 44

Individualized Health Background Method Application Summary

Step I-C: Model for 2-Dimensional Image Dewarping

Prior on the peak-to-landmark indicators

  • Peak-to-landmark Indicators:
  • 1. Zgij ∈ {1, . . . , L}, j = 1, . . . , Jgi (match a “∗” to a “+”), e.g.,

Zgij = 3 means the peak is matched to Landmark 3

  • 2. Constrain Zgi,j−1 ≤ Zgij to prevent reverse matching
  • Bayesian Model for Aligning Peaks to Landmarks
  • Number of observed peaks in lane i, gel g:

Jgi

d

∼ Poisson(Λ) Λ: Cumulative intensity; Controls the total number of peaks

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 18 / 30

slide-45
SLIDE 45

Individualized Health Background Method Application Summary

Step I-C: Model for 2-Dimensional Image Dewarping

Prior on the peak-to-landmark indicators

  • Peak-to-landmark Indicators:
  • 1. Zgij ∈ {1, . . . , L}, j = 1, . . . , Jgi (match a “∗” to a “+”), e.g.,

Zgij = 3 means the peak is matched to Landmark 3

  • 2. Constrain Zgi,j−1 ≤ Zgij to prevent reverse matching
  • Bayesian Model for Aligning Peaks to Landmarks
  • Number of observed peaks in lane i, gel g:

Jgi

d

∼ Poisson(Λ) Λ: Cumulative intensity; Controls the total number of peaks

  • Peak-to-landmark indicators:

(Zgi1, . . . , ZgiJgi) = increasing sort {Z ∗

gi1, . . . , Z ∗ giJgi}

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 18 / 30

slide-46
SLIDE 46

Individualized Health Background Method Application Summary

Step I-C: Model for 2-Dimensional Image Dewarping

Prior on the peak-to-landmark indicators

  • Peak-to-landmark Indicators:
  • 1. Zgij ∈ {1, . . . , L}, j = 1, . . . , Jgi (match a “∗” to a “+”), e.g.,

Zgij = 3 means the peak is matched to Landmark 3

  • 2. Constrain Zgi,j−1 ≤ Zgij to prevent reverse matching
  • Bayesian Model for Aligning Peaks to Landmarks
  • Number of observed peaks in lane i, gel g:

Jgi

d

∼ Poisson(Λ) Λ: Cumulative intensity; Controls the total number of peaks

  • Peak-to-landmark indicators:

(Zgi1, . . . , ZgiJgi) = increasing sort {Z ∗

gi1, . . . , Z ∗ giJgi}

  • Z ∗

gij iid

∼ Categorical

  • {λ∗

ℓ}L ℓ=1

  • Zhenke Wu(zhenkewu@umich.edu)

RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 18 / 30

slide-47
SLIDE 47

Individualized Health Background Method Application Summary

Step I-C: Model for 2-Dimensional Image Dewarping

Prior on the peak-to-landmark indicators

  • Peak-to-landmark Indicators:
  • 1. Zgij ∈ {1, . . . , L}, j = 1, . . . , Jgi (match a “∗” to a “+”), e.g.,

Zgij = 3 means the peak is matched to Landmark 3

  • 2. Constrain Zgi,j−1 ≤ Zgij to prevent reverse matching
  • Bayesian Model for Aligning Peaks to Landmarks
  • Number of observed peaks in lane i, gel g:

Jgi

d

∼ Poisson(Λ) Λ: Cumulative intensity; Controls the total number of peaks

  • Peak-to-landmark indicators:

(Zgi1, . . . , ZgiJgi) = increasing sort {Z ∗

gi1, . . . , Z ∗ giJgi}

  • Z ∗

gij iid

∼ Categorical

  • {λ∗

ℓ}L ℓ=1

  • λ∗

ℓ: Landmark-specific intensity; Independent of g; Hence,

when possible, encourages nearby peaks to be aligned to an identical landmark

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 18 / 30

slide-48
SLIDE 48

Individualized Health Background Method Application Summary

Step I-C: 2-Dimensional Image Dewarping

Gaussian Mixture Model for Noisy Peak Locations “∗”

  • Model the observed peaks Tgij as observations from a L-component

Gaussian mixture, for each candidate landmark ℓ

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 19 / 30

slide-49
SLIDE 49

Individualized Health Background Method Application Summary

Step I-C: 2-Dimensional Image Dewarping

Gaussian Mixture Model for Noisy Peak Locations “∗”

  • Model the observed peaks Tgij as observations from a L-component

Gaussian mixture, for each candidate landmark ℓ

  • We assume

p          (Tgij = t

peak location

, ugi

  • lane

number

) | Zgij = ℓ

matched to landmark ℓ

, Tgi,j−1

nearest left peak location

, Sg

  • warping

function

, σǫ

  • noise

level

         =

  • φ (t; Sg(νℓ, ugi), σǫ) ,

t ∈ Igij(νℓ, A0); 0,

  • therwise,

ℓ = 1, . . . , L, peak j = 1, . . . , Jgi, lane i = 1, . . . , Ng, gel g = 1, . . . , G.

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 19 / 30

slide-50
SLIDE 50

Individualized Health Background Method Application Summary

Step I-C: 2-Dimensional Image Dewarping

Gaussian Mixture Model for Noisy Peak Locations “∗”

  • Model the observed peaks Tgij as observations from a L-component

Gaussian mixture, for each candidate landmark ℓ

  • We assume

p          (Tgij = t

peak location

, ugi

  • lane

number

) | Zgij = ℓ

matched to landmark ℓ

, Tgi,j−1

nearest left peak location

, Sg

  • warping

function

, σǫ

  • noise

level

         =

  • φ (t; Sg(νℓ, ugi), σǫ) ,

t ∈ Igij(νℓ, A0); 0,

  • therwise,

ℓ = 1, . . . , L, peak j = 1, . . . , Jgi, lane i = 1, . . . , Ng, gel g = 1, . . . , G.

  • φ(·; a, b): Gaussian density with mean a and standard deviation b.

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 19 / 30

slide-51
SLIDE 51

Individualized Health Background Method Application Summary

Step I-C: 2-Dimensional Image Dewarping

Gaussian Mixture Model for Noisy Peak Locations “∗”

  • Model the observed peaks Tgij as observations from a L-component

Gaussian mixture, for each candidate landmark ℓ

  • We assume

p          (Tgij = t

peak location

, ugi

  • lane

number

) | Zgij = ℓ

matched to landmark ℓ

, Tgi,j−1

nearest left peak location

, Sg

  • warping

function

, σǫ

  • noise

level

         =

  • φ (t; Sg(νℓ, ugi), σǫ) ,

t ∈ Igij(νℓ, A0); 0,

  • therwise,

ℓ = 1, . . . , L, peak j = 1, . . . , Jgi, lane i = 1, . . . , Ng, gel g = 1, . . . , G.

  • Sg: (νℓ, ugi) → Sg(νℓ, ui), unknown, smooth bivariate function for the

spatial deformation

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 19 / 30

slide-52
SLIDE 52

Individualized Health Background Method Application Summary

Step I-C: 2-Dimensional Image Dewarping

Gaussian Mixture Model for Noisy Peak Locations “∗”

  • Model the observed peaks Tgij as observations from a L-component

Gaussian mixture, for each candidate landmark ℓ

  • We assume

p          (Tgij = t

peak location

, ugi

  • lane

number

) | Zgij = ℓ

matched to landmark ℓ

, Tgi,j−1

nearest left peak location

, Sg

  • warping

function

, σǫ

  • noise

level

         =

  • φ (t; Sg(νℓ, ugi), σǫ) ,

t ∈ Igij(νℓ, A0); 0,

  • therwise,

ℓ = 1, . . . , L, peak j = 1, . . . , Jgi, lane i = 1, . . . , Ng, gel g = 1, . . . , G.

  • The set Igij(νℓ, A0) ∆

= {t : |t − νℓ| < A0 and t > Tgi,j−1} assumes a peak appears within distance A0 from its true landmark

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 19 / 30

slide-53
SLIDE 53

Individualized Health Background Method Application Summary

Step I-C: 2-Dimensional Image Dewarping

Gaussian Mixture Model for Noisy Peak Locations “∗”

  • Model the observed peaks Tgij as observations from a L-component

Gaussian mixture, for each candidate landmark ℓ

  • We assume

p          (Tgij = t

peak location

, ugi

  • lane

number

) | Zgij = ℓ

matched to landmark ℓ

, Tgi,j−1

nearest left peak location

, Sg

  • warping

function

, σǫ

  • noise

level

         =

  • φ (t; Sg(νℓ, ugi), σǫ) ,

t ∈ Igij(νℓ, A0); 0,

  • therwise,

ℓ = 1, . . . , L, peak j = 1, . . . , Jgi, lane i = 1, . . . , Ng, gel g = 1, . . . , G.

  • Let Pg be the peaks for gel g; let P collect all the peaks

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 19 / 30

slide-54
SLIDE 54

Individualized Health Background Method Application Summary

Step I-C: 2-Dimensional Image Dewarping

Warping Function by Tensor Product Basis Expansion

  • We assume the warping function

Sg(ν, u) =

Tν

  • s=1

Tu

  • t=1

βgstBg1s(ν)Bg2t(u),

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 20 / 30

slide-55
SLIDE 55

Individualized Health Background Method Application Summary

Step I-C: 2-Dimensional Image Dewarping

Warping Function by Tensor Product Basis Expansion

  • We assume the warping function

Sg(ν, u) =

Tν

  • s=1

Tu

  • t=1

βgstBg1s(ν)Bg2t(u),

  • Bg1s(·) and Bg2t(·): the s-th and t-th cubic B-spline basis

along the two coordinate directions, respectively

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 20 / 30

slide-56
SLIDE 56

Individualized Health Background Method Application Summary

Step I-C: 2-Dimensional Image Dewarping

Warping Function by Tensor Product Basis Expansion

  • We assume the warping function

Sg(ν, u) =

Tν

  • s=1

Tu

  • t=1

βgstBg1s(ν)Bg2t(u),

  • Bg1s(·) and Bg2t(·): the s-th and t-th cubic B-spline basis

along the two coordinate directions, respectively

  • {βgst}: the set of coefficients to be estimated

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 20 / 30

slide-57
SLIDE 57

Individualized Health Background Method Application Summary

Step I-C: 2-Dimensional Image Dewarping

Warping Function by Tensor Product Basis Expansion

  • We assume the warping function

Sg(ν, u) =

Tν

  • s=1

Tu

  • t=1

βgstBg1s(ν)Bg2t(u),

  • Bg1s(·) and Bg2t(·): the s-th and t-th cubic B-spline basis

along the two coordinate directions, respectively

  • {βgst}: the set of coefficients to be estimated
  • Implementing Warping Function Constraints and Priors

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 20 / 30

slide-58
SLIDE 58

Individualized Health Background Method Application Summary

Step I-C: 2-Dimensional Image Dewarping

Warping Function by Tensor Product Basis Expansion

  • We assume the warping function

Sg(ν, u) =

Tν

  • s=1

Tu

  • t=1

βgstBg1s(ν)Bg2t(u),

  • Bg1s(·) and Bg2t(·): the s-th and t-th cubic B-spline basis

along the two coordinate directions, respectively

  • {βgst}: the set of coefficients to be estimated
  • Implementing Warping Function Constraints and Priors
  • Boundary constraint: Sg(ν0, u) = ν0, Sg(νL+1, u) = νL+1

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 20 / 30

slide-59
SLIDE 59

Individualized Health Background Method Application Summary

Step I-C: 2-Dimensional Image Dewarping

Warping Function by Tensor Product Basis Expansion

  • We assume the warping function

Sg(ν, u) =

Tν

  • s=1

Tu

  • t=1

βgstBg1s(ν)Bg2t(u),

  • Bg1s(·) and Bg2t(·): the s-th and t-th cubic B-spline basis

along the two coordinate directions, respectively

  • {βgst}: the set of coefficients to be estimated
  • Implementing Warping Function Constraints and Priors
  • Boundary constraint: Sg(ν0, u) = ν0, Sg(νL+1, u) = νL+1
  • Monotonic constraint:

ν0 ≤ Sg(ν, u) < Sg(ν′, u ≤ νL+1, ∀ν < ν′, ∀u

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 20 / 30

slide-60
SLIDE 60

Individualized Health Background Method Application Summary

Step I-C: 2-Dimensional Image Dewarping

Warping Function by Tensor Product Basis Expansion

  • We assume the warping function

Sg(ν, u) =

Tν

  • s=1

Tu

  • t=1

βgstBg1s(ν)Bg2t(u),

  • Bg1s(·) and Bg2t(·): the s-th and t-th cubic B-spline basis

along the two coordinate directions, respectively

  • {βgst}: the set of coefficients to be estimated
  • Implementing Warping Function Constraints and Priors
  • Boundary constraint: Sg(ν0, u) = ν0, Sg(νL+1, u) = νL+1
  • Monotonic constraint:

ν0 ≤ Sg(ν, u) < Sg(ν′, u ≤ νL+1, ∀ν < ν′, ∀u

  • Both constraints above can be implemented via constraints on

{βgst}

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 20 / 30

slide-61
SLIDE 61

Individualized Health Background Method Application Summary

Step I-C: 2-Dimensional Image Dewarping

Warping Function by Tensor Product Basis Expansion

  • We assume the warping function

Sg(ν, u) =

Tν

  • s=1

Tu

  • t=1

βgstBg1s(ν)Bg2t(u),

  • Bg1s(·) and Bg2t(·): the s-th and t-th cubic B-spline basis

along the two coordinate directions, respectively

  • {βgst}: the set of coefficients to be estimated
  • Implementing Warping Function Constraints and Priors
  • Boundary constraint: Sg(ν0, u) = ν0, Sg(νL+1, u) = νL+1
  • Monotonic constraint:

ν0 ≤ Sg(ν, u) < Sg(ν′, u ≤ νL+1, ∀ν < ν′, ∀u

  • Both constraints above can be implemented via constraints on

{βgst}

  • Smoothness: Bayesian penalized-splines to make adjacent

{βgst} similar

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 20 / 30

slide-62
SLIDE 62

Individualized Health Background Method Application Summary

Step I-C: 2-Dimensional Image Dewarping

Warping Function by Tensor Product Basis Expansion

  • We assume the warping function

Sg(ν, u) =

Tν

  • s=1

Tu

  • t=1

βgstBg1s(ν)Bg2t(u),

  • Bg1s(·) and Bg2t(·): the s-th and t-th cubic B-spline basis

along the two coordinate directions, respectively

  • {βgst}: the set of coefficients to be estimated
  • Implementing Warping Function Constraints and Priors
  • Boundary constraint: Sg(ν0, u) = ν0, Sg(νL+1, u) = νL+1
  • Monotonic constraint:

ν0 ≤ Sg(ν, u) < Sg(ν′, u ≤ νL+1, ∀ν < ν′, ∀u

  • Both constraints above can be implemented via constraints on

{βgst}

  • Smoothness: Bayesian penalized-splines to make adjacent

{βgst} similar

  • Vary by gel: Sg(νℓ, u)

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 20 / 30

slide-63
SLIDE 63

Individualized Health Background Method Application Summary

Step I-C: A Mathematical Model for Warping

Estimate the warping, then reverse

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 21 / 30

slide-64
SLIDE 64

Individualized Health Background Method Application Summary

Step I-C: Goal of 2-Dimensional Image De-warping

The posterior distribution [Z | P] Recall:

  • Z: the collection of peak-to-landmark indicators

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 22 / 30

slide-65
SLIDE 65

Individualized Health Background Method Application Summary

Step I-C: Goal of 2-Dimensional Image De-warping

The posterior distribution [Z | P] Recall:

  • Z: the collection of peak-to-landmark indicators
  • P: the collection of all the observed peaks

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 22 / 30

slide-66
SLIDE 66

Individualized Health Background Method Application Summary

Step I-C: Posterior Inference of the De-Warping

Joint distribution [P, Z]:

  • Goal: Joint distribution [P, Z](data+unknowns) → Posterior

distribution [Z | P] (unknown given data)

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 23 / 30

slide-67
SLIDE 67

Individualized Health Background Method Application Summary

Step I-C: Posterior Inference of the De-Warping

Joint distribution [P, Z]:

  • Goal: Joint distribution [P, Z](data+unknowns) → Posterior

distribution [Z | P] (unknown given data)

  • Tool: Markov chain Monte Carlo (MCMC)

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 23 / 30

slide-68
SLIDE 68

Individualized Health Background Method Application Summary

Step I-C: Posterior Inference of the De-Warping

Joint distribution [P, Z]:

  • Goal: Joint distribution [P, Z](data+unknowns) → Posterior

distribution [Z | P] (unknown given data)

  • Tool: Markov chain Monte Carlo (MCMC)
  • Idea: Simulate samples from the joint posterior distribution of

the unknowns given the data;

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 23 / 30

slide-69
SLIDE 69

Individualized Health Background Method Application Summary

Step I-C: Posterior Inference of the De-Warping

Joint distribution [P, Z]:

  • Goal: Joint distribution [P, Z](data+unknowns) → Posterior

distribution [Z | P] (unknown given data)

  • Tool: Markov chain Monte Carlo (MCMC)
  • Idea: Simulate samples from the joint posterior distribution of

the unknowns given the data; Then use the samples to do posterior inference for any functions of the unknowns

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 23 / 30

slide-70
SLIDE 70

Individualized Health Background Method Application Summary

Step I-C: Align the peaks – Result

Animation; “∆” for signature; “•” for the observed peaks

(Please Click the Image for Animation)

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 24 / 30

slide-71
SLIDE 71

Individualized Health Background Method Application Summary

Step I-C: Aligned High-Frequency Intensity Data

Before

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 25 / 30

slide-72
SLIDE 72

Individualized Health Background Method Application Summary

Step I-C: Aligned High-Frequency Intensity Data

Before

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 25 / 30

slide-73
SLIDE 73

Individualized Health Background Method Application Summary

Step I-C: Aligned High-Frequency Intensity Data

Before After

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 25 / 30

slide-74
SLIDE 74

Individualized Health Background Method Application Summary

Step I-C: Aligned High-Frequency Intensity Data

Before note the curvatures are removed − − − − − − − − − − − − − − − − − − − − − − − − − → After

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 25 / 30

slide-75
SLIDE 75

Individualized Health Background Method Application Summary

Data

Scleroderma

  • Long-term clinical objective: find autoantibody signature that

subsets autoimmune disease patients into groups with more homogeneous phenotypes and trajectories

  • Sera from well-characterized patients with scleroderma and an

associated cancer from Johns Hopkins Scleroderma Center database

  • Data
  • 1. Known clustering: two replicate GEA experiments on 20

samples

  • 2. Unknown clustering: non-replicate GEA experiment on 80

samples

  • Steps:
  • 1. Pre-processing
  • 2. Clustering (into 2, 3, . . .,N groups) based on the pre-processed

high-frequency intensity data (hierarchical clustering here)

  • 3. Evaluate the separation of the obtained clusters and compare

them to the truth (known in the replicate experiment)

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 26 / 30

slide-76
SLIDE 76

Individualized Health Background Method Application Summary

Pre-processing Improves the Accuracy of Cluster Estimation

Data with technical replicates; 20 samples, long- and short- exposures

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 27 / 30

slide-77
SLIDE 77

Individualized Health Background Method Application Summary

Pre-processing Improves the Separation of Clusters

Data without Replicates; Hierarchical Clustering; Pre-processed vs Non-Pre-processed

  • Distance: Correlation-based distance; complete linkage
  • Interpretation: adjacent terminal nodes in the tree → similar in

AutoAntibody signatures

  • Uncertainty: confidence levels by multiscale boostrapping (red numbers;
  • nes > 95 are shown in red boxes; a numbering of the subtrees is shown in

blue)

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 28 / 30

slide-78
SLIDE 78

Individualized Health Background Method Application Summary

Pre-processing Improves the Separation of Clusters

Data without Replicates; Hierarchical Clustering; Pre-processed vs Non-Pre-processed

  • Distance: Correlation-based distance; complete linkage
  • Interpretation: adjacent terminal nodes in the tree → similar in

AutoAntibody signatures

  • Uncertainty: confidence levels by multiscale boostrapping (red numbers;
  • nes > 95 are shown in red boxes; a numbering of the subtrees is shown in

blue)

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 28 / 30

slide-79
SLIDE 79

Individualized Health Background Method Application Summary

Summary

  • Problem: Human recognition of autoantibody patterns and hence

clustering becomes more difficult when patterns are composite and

  • n multiple gels
  • Method: Novel automated algorithms that
  • 1. Estimate autoantibody signatures
  • 2. The pre-processed data (Step I) can be the input of many subgroup

discovery methods (Step II) including hierarchical clustering, latent class models and factor analyses

  • 3. Improves the accuracy of subgroup discovery
  • Free publicly available open-source software:

https://github.com/zhenkewu/spotgear

  • Manuscript: Wu, Casciola-Rosen, Shah, Rosen, Zeger (2017).

http://biorxiv.org/content/early/2017/04/21/128199

  • Ongoing work: novel Bayesian clustering model to find disease

subsets; Based on the biology that autoantibodies recognize protein complexes.

Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 29 / 30

slide-80
SLIDE 80

Individualized Health Background Method Application Summary

Thank You!

Funding Patient-Centered Outcome Research Institute [PCORI ME-1408-20318] Hopkins Individualized Health Initiative Some References (More at: zhenkewu.com)

  • Wu Z, Casciola-Rosen L, Shah AA, Rosen A, Zeger SL (2017+).

Estimating AutoAntibody Signatures to Detect Autoimmune Disease Patient Subsets. Minor Revision for Biostatistics. http://biorxiv.org/content/early/2017/04/18/128199.

  • Wu Z, Deloria-Knoll M, Hammitt LL, and Zeger SL, for the PERCH Core Team (2015).

Partially Latent Class Models (pLCM) for Case-Control Studies of Childhood Pneumonia Etiology. Journal of the Royal Statistical Society: Series C (Applied Statistics). 65:97-114.

  • Wu Z, Deloria-Knoll M and Zeger SL (2016a).

Nested Partially-Latent Class Models for Estimating Disease Etiology from Case-Control Data. Biostatistics, 18 (2): 200-213. doi:10.1093/biostatistics/kxw037. Zhenke Wu(zhenkewu@umich.edu) RBras62, UFLA, Lavras, MG, Brazil 27 July 2017 30 / 30