Introduction in the analysis of cell-based functional...

Florian HahneDepartment of Molecular Genome Analysis

German Cancer Research Center Heidelberg

Introduction in the analysis of cell-based functional assay using Bioconductor

CSAMA Bressanone 2006

Candidate gene sets from microarray

studies: dozens…hundreds

Capacity of detailed in-vivo functional studies:

one…few

How to close the gap?

How to separate a flood of ‘significant’secondary effects from causally relevant ones?

Why do we need functional assays?

• means to monitor effect of perturbation

expression or activation state of key regulatory proteins (plate reader, FACS, automated microscope)

The design: manipulate gene expression/protein function

• means to monitor perturbation (beneficial but not mandatory)

expression of fluorescence protein tag

• system to willfully manipulate expression level of certain genes in cells

up regulation (transfection of expression vectors)

down regulation (RNA interference, small compounds)

Features of cell-based assays: reporter system

Moffat et al. Nature Reviews Molecular Cell Biology 7, 177–187 (March 2006)

Features of cell-based assays

• What is the right assay for the question asked?In principle, any cellular function can be probed

• A cell is a complex system assay needs optimization

• Not all cells are the same choice of appropriate cell system

• Different levels of complexity:cell populations, individual cells, cell components

• Different technologiesplate readers, flow cytometry, microscopy

• Large amounts of data (genome-wide)information services, visualization, data management

functional assayphenotypes

Differentially regulatedgenes

~ 70 280 genes

RIP/IMDpathways

RIP

Tak1

IKK

Rel

R

Targets

Michael Boutros

Is differential expression a good predictor for ’signaling’ function?

Most pathway targets are not required for pathway function

functional assayphenotypes

Differentially regulatedgenes

~ 70 280 genes

3

RIP/IMDpathways

RIP

Tak1

IKK

Rel

R

Targets

Michael Boutros

Is differential expression a good predictor for ’signaling’ function?

RNAi is a post-transcriptional gene-silencing process...

...that can be applied in a high-throughput fashion on the cell or

organism scale.

RNA interference (RNAi)

T7

Precursor dsRNA

siRNAs

Degradation of target message

C. elegans Drosophila Mammals

Injection and soaking

Feeding bacteria

Worms

Bathing

> 200bp> 200bp 21bp

Cell culture

dsRNA dsRNA siRNAE. coli

DICER

RNAi experiments in different organisms

Moffat et al. Nature Reviews Molecular Cell Biology 7, 177–187 (March 2006)

Large scale RNAi screens

• R/Bioconductor package• Systematic analysis and documentation of cell-based HTS assays by RNAi or other type of perturbation libraries• Step-by-step analysis, from raw data files to the annotated hit list• Prerequisite for multiple screen comparisons (standardization)• Audit trail of experiment QA• The whole experiment is contained in one object

cellHTS package

Computational statistics using bioinformatic tools

Hit identification

Candidates

Gene Ontology

Expression DB

Protein DB

Raw data

Integrate information from external DB

Hit validation by secondary assays

cellHTS

cellHTSpackage

package

Screening work-flow

cellHTSpackage

A typical cell-based HTS assay

Import the raw data files

Plate-wise quality control

Data preprocessing

Experiment-wise quality control

Ranking of phenotypes (hit list)

HTML quality reports

Work flow

cell

num

ber

plate plots as graphical representation of experimental entities

• false color coding for concise display of numeric outcomes from statistical analyses

visualization of results

quantitative

Visualization: plate plots (in package prada)




Visualization: plate plots (in package prada)

qualitative

• per plate quality assessment• Dynamic range;

• Scatter plot between replicates and correlation coefficient;

• Distribution of the intensity values for each replicate;

• “Plate plots” for the replicate measurements and for the standard deviation between replicate measurements

• per experiment quality assessment• Boxplots for each replicate grouped by plate;

• Distribution of the signal in the control wells

• whole screen visualization

Main features

Quality reports in the form of HTML pages

cellHTS package

• Plate effects (plate-to-plate variations)

• Edge or spatial effects within the plate (well-to-well variations)

Remove systematic biases and variations, while keeping the biological relevant information

Assay formats, pipetting delivery, robotic failures, differences in compound concentrations due to

evaporation of solvent, potency differences across compounds, systematic across-plate biases, within-plate

spatial biases, ...

Data preprocessing

• Percent of control:

• Normalized percent inhibition:

• z score:

Plate effects

kth wellith plate

Plate effects & edge effects

• B score:

100×positivei

ki'ki µ

x=x

negativei

positivei

kipositive

i'ki µµ

xµ=x−−

i

iki'ki σ

µx=x −

rth rowcth column

ith plate

( )i

ciriirci'rci MAD

CRx=xˆˆˆ ++− µ

Data preprocessing

• Scale the plates by a plate-specific factor:

• Based on the intensities of the controls

• Based on the intensities of the samples

• Robust location estimators that take into account the assignment of RNAi reagents to the plates (random or non-random)

Plate effects

kth wellith plate

Data preprocessing

Plate effects – plate median scaling

kth wellith plate

Data preprocessing

Data preprocessing: plate effects due to library design

raw data

Plate 26

proteasome subunits or components;

ATP/GTP-binding site motifs

ribosomal proteins

like-Sm nucleoproteins and ribosomal proteins

Data preprocessing : plate effects due to library design

• Consider the shorth of the distribution of intensities in each plate as the per-plate scaling factor.

Plate effects & the siRNA library

Data preprocessing : plate effects due to library design

Data preprocessing: plate effects due to library design

before normalization after normalization

Median intensity across plates (preprocessed data)

Data preprocessing: edge effects

Data preprocessing: edge effects

Look at the sample variances based on the normalized replicate values

For some cases data transformation can stabilize variance

Data preprocessing: data transformation

Data Transformation: viability screen

Data transformation: treatment screen

Toll and Imd screens in D. melanogaster cells

• Two reporters:

• Firefly luciferase (F) for pathway activity

• Renilla luciferase (R) for growth and viability

• How to combine the intensities of the two reporters?

Two-color data

Toll screen

viability

path

way

activ

ity

FACS (fluorescence activated cell sorting)

light scatter detector

Fluorescence detector(PMT3, PMT4 etc.)

Laser

• measures fluorescence intensities as well as morphological parameters on the basis of light emission

• offers single cell resolution

• robust, reliable, variable

ORF

ORF

ORF

attB1attB2

attB1attB2

attB1attB2

ORF

attL1

attL2

entryclone

ORF

ORF

ORF

attB1attB2

attB1attB2

attB1attB2

PCR amplification

ORF

attL1

attL2

entryclone

Full coding cDNA clone

ORF cloning: The Gateway™ System

N ORF YFP CORFYFPN CN-terminal tag C-terminal tag

- package prada

package prada contains functionalities for analysis of data derived from cell based assays with a strong focus on flow cytometry data

modular framework

• data preprocessing• data visualization• data integration and management

for statistical inference and modeling general purpose tools can be used

• linear models • local regression• hypothesis testing

• FCS 3.0 files- standardized storage format for FACS data- contains fluorescence values in data segment, wealth of meta

data in text segment- can be imported into R (function readFCS)

Data import and maintenance (object orientation)

• cytoFrameR internal representation of data from one FCS file

- raw data matrix

- list of meta data• cytoSet

R internal representation of data from several FCS files (e.g. one 96 well plate)

generic functions, class methods

Gating

Gate: Selections of subpopulations of cells with respect to one or several measurement parameters.

Objects of class gate and gateSet

interactive drawing of gates based on two-dimensional scatter plots

can be assigned to cytoFrames

Gating

G1 + G2

G1 – G2

G1 ∩ G2

G2

G2

G2

combination of gates:

G1

G1

G1

distinction on basis of morphological properties

strong variation between experiments

dynamic determination

cell size

gran

ular

ityData pre-processing: FSC vs. SSC plot

Data pre-processing: finding the main population

assumption:bivariate normal distribution

robust fitting

discarding cells that do not lie within some given boundary of this distribution

=density ofdistribution

= discarded

X =midpoint ofdistribution

Data pre-processing: finding the main population

=density ofdistribution

= discarded

X =midpoint ofdistribution

shape and location of main distribution can be used for quality control

assumption:bivariate normal distribution

robust fitting

discarding cells that do not lie within some given boundary of this distribution

Understanding FACS data

parameter 1(perturbation)

para

met

er 2

(phe

noty

pe)



para

met

er 2

(phe

noty

pe)activation



para

met

er 2

(phe

noty

pe)inhibition

Dealing with correlations

cell size correlates with fluorescent intensities

(FL1)

(FL4)

specifictotal xsx ++= βα

induces spurious correlationsin the data

s: cell size (FSC) xtotal : measured fluorescencexspecific: actual fluorescence emitted by dye

different responses for different assays

• discrete response: on/off mechanism(e.g. apoptosis, proliferation)

over expression

effe

ct

over expression

effe

cttheory FACS

• continuous response: concentration dependent(e.g. MAP kinase)

over expression

effe

ct

over expression

effe

ct

theory FACS

Statistical analysis: mode of response

• robust fitting of smoothed local regression function:y: response (phenotype)x: perturbation signalm: smooth function

: robust estimator of m at point x0

• z-score as dimension less measure of effect:ratio of estimated slope δ at point x0 and assay-widescale parameter δ0

z = 18.1 z = 0.4 z = -40.2

t* t* t*

Statistical analysis: continuous response

( )( )0xmxmy′=

+=)δ

ε

0δδ

=z

)(ˆ 0xm′

Fisher’s exact test

Statistical analysis: discrete response

non perturbedpositive

(a)

non perturbednegative

(b)

perturbednegative

(d)

perturbedpositive

(c)ph

enot

yype

perturbation

, p valueeffect size significance

2

1

rrratioodds =

11

1 ++

=bar

11

2 ++

=dcr

Statistical analysis: discrete response

no effect activator

17 440

9556 3247

58 64

6010 5321

-log(odds ratio) = 0.09(p = 0.24)

-log(odds ratio) = 4.33(p = 2.2e-16)




• HTML image map allows for hyper linking to include further information for each well

Visualization: plate plots

additionalinformation






replicates






anything…

Package rflowcyt

• Also deals with flow cytometry data

• Focus more on individual FACS measurements and quality control

• slightly different object model but essentially the same conceptconversion functions from and to cytoFrames

• Quality assessment tools

Package rflowcyt

box plotcontour plot

ECDF density plot summary plot

Control phenotype

What the data look like...

Nuclear phenotype Cytokinesis

Mitotic arrest

Multipolar spindles

Tubulin elongation

CONTROL

What we try to find...

dm <- sqrt(distMap(seg))res <- objectCount(dm, gray[,,4], 100, 40)

index x y size intensity[1,] 1 304 95 1065 221.95894[2,] 1 140 186 680 141.64695[3,] 1 222 217 786 178.99816[4,] 1 0 170 274 61.27550[5,] 1 336 139 800 148.25224[6,] 1 212 91 696 213.69449[7,] 1 290 267 664 150.84269[8,] 1 107 101 1102 245.86509[9,] 1 257 0 372 83.80994

Image processing and analysis: Bioconductor

package EBImage(ImageMagick & others)

Computational statistics on vector data:

clustering, classification, hypothesis testing

R

Phenotypes, gene functions

ReproducibilityEvolution of code

ParallelizationDon't reinvent the wheel, stand on the

shoulders of giants

Features

Supports a variety of 2D (about 40) and some 3D image formats (TIFF) in read/write

mode, supports local file system and network protocols (HTTP, FTP)

Image objects are based on native R arrays thus supporting all functions

available for arrays as well as giving speed in operating with images in R

Majority of code is C/C++ for high performance

I/O is based on ImageMagick::Magick++ (C++) library available for a number of

platforms and operating systems (including both Windows and Linux)

Effective memory management to enable operations on very large data sets

Majority of ImageMagick:Magick++ 2D image processing filters:

threshold, blur, noise removal, edge, sharpen, unsharp mask etc

Distance Map filter, Object Counting algorithms

Manipulating Image Data

im1 <- read.image(“im01.jpg”); im2 <- read.image(“im02.jpg”)

# addition of two images, combining features of both in oneim3 <- im1 + im2

# subtraction of images – image differenceim4 <- im1 – im2

# multiplication – amplification of common features and removal of differencesim5 <- im1 * im2

# scaling of dataim6 <- im1 * 2

# extending contrast of dark regionsim7 <- sqrt(im1)

# cropping images and subscriptingim8 <- im1[100:200, 80:180]im9 <- im1[100:200, ]

# conditional replacement of image data – thresholdingim8[im8 > 0.5] <- 1.0

# data of one image is modified based on condition from another oneim1[im2 <= 0.2] <- 0.0

# conversions between colour modes and summation of RGB imagesrgb <- toRed(im1) + toGreen(im2); gray <- toGray(rgb)

addition

subtraction

multiplication

sqrt(im)

im[im>0.4]=1

im[..]

Image filters

original imageim <- read.image(..)

normalized in [0..1]im <- normalize(im)

false colour, 2-channelsim2 <- toRed(im) +

toGreen(im1)

adaptive thresholdingseg <- thresh(im, 20,

20, 400, TRUE)

skeletonsk <- edge(dm, 1)

edge filtered <- edge(im, 1)

Other filtersenhancements: blur, despeckle, enhance, medianFilter, gaussianFilter, redNoise, sharpen,spread, unsharpMasksegmentation: segment

colour: contrast, equalize,colorGamma, mod, shadetransformations: rotate, sample.image, scale.image

Distance mapsdm <- distMap(seg)display(normalize(dm))

Object count & indexingres <- objectCount(dm)

Identifying cells

Object marking

All filters are implemented in C++

for high performance

Acknowledgements

• EBI:

Wolfgang Huber

Ligia Bras

Oleg Sklyar

• DKFZ:

Stefan Wiemann

Dorit Arlt

Meher Majety

Mamatha Sauerman

Michael Boutros

Florian Fuchs

Viola Gesellchen

Dierk Ingelfinger

David Kuttenkeuler

Sandra Steinbrink

• FHCRC

Robert Gentleman

Nolwenn LeMeur

Seth Falcon

Introduction in the analysis of cell-based functional...

Documents

Transcript of Introduction in the analysis of cell-based functional...