Tiered Computation Raymond Ng CIO, Proof Centre Acting Head, - - PowerPoint PPT Presentation

Tiered Computation

Raymond Ng CIO, Proof Centre Acting Head, Department of Computer Science, UBC

5/4/2011 2

BiT Biomarker Discovery Strategy

“Omics” Tools and Approaches

BIOMARKER BIOLIBRARY Blood Urine Tissue

METABOLOMICS

Serum and Urine

U of Alberta Metabolomics Platform, Edmonton, AB

NMR & Mass Spec Analysis

TRANSCRIPTOMICS

PAXgene Whole Blood

Microarray Core Laboratory, Children’s Hospital, LA, CA

Affymetrix Microarray Analysis RNA Extraction

PROTEOMICS

Nascent Plasma Depleted Plasma Bound to Column

Albumin

Plasma

UVic-Genome BC Proteomics Platform, Victoria, BC

ABI 4800 iTRAQ Analysis Plasma Depletion

QA/QC – All sample collection and processing is done to SOP

5/4/2011 3 3

Importance of Data Cleansing and Pre-processing

• A. Clinical: “Detecting potential labeling errors in

microarrays by data perturbation,’’ Bioinformatics 2006

(Malossini, Blanzieri)

• B. mRNA: “MDQC: a new quality assessment method

for microarrays based on quality control reports,”

Bioinformatics 2007 (Cohen-Freue, Hollander et al.)

• C. DNA: “Modelling Recurrent DNA Copy Number

Alterations in array CGH Data,” Bioinformatics 2006, 2007

(Shah, Murphy, Lam)

5/4/2011 4 4

50 100 150 200 500 1000 1500 21-4

Sample Quality

50 100 150 200 100 200 300 400 Sample 21-4 17-6 25-5 302-7

Chip Quality

50 100 150 200 5 10 15 Sample 317-10 13-2 13-3 13-4 13-5 13-6 19-1 320-1

RNA Quality

Microarray Quality Control Assessment Tool

Finding “Needles in a Haystack”

5

Pre-filtering

54,000 Probe Sets 2,000 Proteins/Metabolites

< 10,000 Probe Sets < 100 Proteins/Metabolites

I. Remove features with small variations across all samples (rejection

• r otherwise)

54,000 Probe Sets 2,000 Proteins/Metabolites

~ 100-500 Genes/Proteins/Metabolites/ Clinical Variables

< 10,000 Probe Sets < 100 Proteins/Metabolites

“Needles in the Haystack” (cont)

6

Ranking and Filtering

• II. (Univariate) Rank each

individual feature on how well it discriminates the rejection samples from non-rejection samples

• III. (Multi-variate) Rank groups
• f features together on their

joint discrimination power, taking correlation into account

BIOMARKER PANEL

INTERNALLY VALIDATED BIOMARKER PANEL

54,000 Probe Sets 2,000 Proteins/Metabolites ~ 100-500 Genes/Proteins/Metabolites/ Clinical Variables < 10,000 Probe Sets < 100 Proteins/Metabolites

“Needles in the Haystack” (cont)

7

Panel Selection, Model Building

• IV. Select features to be

included in the panel, possibly assigning different weights to different features

Rich Space for Choices

8

Pre-filtering (remove

probe-sets with low variability) 1) k samples above absolute threshold 2) First half using inter-quartile range 3) First half using empirical central mass range

Uni-variate ranking

(FDR-based; per probe-set) 1) Maximum of LIMMA, robust LIMMA and SAM 2) LIMMA 3) Robust LIMMA

Uni-variate filtering

(per probe set) 1) FDR cut-off (FDR<0.01) 2) Size cut-off: Top 50 probe-sets 3) Combination rule: FDR<0.05 but at least 50 and at most 500 probe sets

Multi-variate ranking (optional)

1) Stepwise Discriminant Analysis 2) SVM-based ranking (one step) 3) Recursive Feature Elimination (multi-step) 4) Elastic Net-based (coefficients)

Multi-variate filtering (optional)

1) Significance of improvement cut-off 2) Top 50 (as returned by multi-variate ranking) 3) Non-zero coefficients (Elastic Net)

Classifier Generation

1) Linear Discriminant Analysis 2) Support Vector Machine 3) Random Forest 4) Elastic Net 5) Logistic regression