Statistical Aspects of the Development and Validation of Predictive Classifiers for High Dimensional...
-
Upload
ralph-harmon -
Category
Documents
-
view
216 -
download
0
Transcript of Statistical Aspects of the Development and Validation of Predictive Classifiers for High Dimensional...
Statistical Aspects of the Development and Validation of Predictive Classifiers for High
Dimensional Data
Richard Simon, D.Sc.Chief, Biometric Research Branch
National Cancer InstituteLinus.nci.nih.gov/brb
BRB Websitehttp://linus.nci.nih.gov/brb
• Powerpoint presentations and audio files
• Reprints & Technical Reports
• BRB-ArrayTools software
• BRB-ArrayTools Data Archive
• Sample Size Planning for Targeted Clinical Trials
DNA Microarray Gene Expression Assay
• Extract mRNA from cells of interest– Each mRNA molecule was transcribed from a single gene and it has a
linear structure complementary to that gene– 1 mRNA molecule is translated into one protein molecule
• Reverse transcribe mRNA to cDNA introducing a fluorescently labeled dye to each molecule
• Distribute the cDNA sample to a solid surface containing “probes” of DNA representing all “genes”– the probes are arranged in a grid on the surface with each gene having
a known address• Let the molecules from the sample hybridize with the probes for the
corresponding genes• Wash off excess sample and illuminate surface with laser with
frequency corresponding to the dye• Measure intensity of fluorescence over each probe
Resulting Data
• Intensity over a probe is approximately proportional to abundance of mRNA molecules in the sample for the gene corresponding to the probe
• 40,000 variables measured for each specimen
Good Microarray Studies Have Clear Objectives
• Class Comparison (Gene Finding)– Find genes whose expression differs among predetermined classes– Tumor versus normal– After infection of cells by virus versus before infection
• Class Prediction– Prediction of predetermined class using gene expression profile– Predicting whether a patient will or will not respond to a given
treatment• Class Discovery
– Discover clusters of specimens having similar expression profiles– Find genes that are in the same biochemical pathways
Class Comparison and Class Prediction
• Not clustering problems
• Supervised methods
Components of Class Prediction
• Feature (gene) selection– Which genes will be included in the model
• Select model type – E.g. Diagonal linear discriminant analysis,
Nearest-Neighbor, …
• Fitting parameters (regression coefficients) for model– Selecting value of tuning parameters
Simple Gene Selection
• Select genes that are differentially expressed among the classes at a significance level (e.g. 0.01) – The level is a tuning parameter– Number of false discoveries is not of direct relevance for
prediction
• For prediction it is usually more serious to exclude an informative variable than to include some noise variables
Optimal significance level cutoffs for gene selection. 50 differentially expressed genes
out of 22,000 genes on the microarrays 2δ/σ n=10 n=30 n=50
1 0.167 0.003 0.00068
1.25 0.085 0.0011 0.00035
1.5 0.045 0.00063 0.00016
1.75 0.026 0.00036 0.00006
2 0.015 0.0002 0.00002
Complex Gene Selection
• Select a set of genes which together give most accurate predictions– Genetic algorithms
• Little evidence that complex feature selection is useful in microarray problems
Linear Classifiers for Two Classes
( )
vector of log ratios or log signals
features (genes) included in model
weight for i'th feature
decision boundary ( ) > or < d
i ii F
i
l x w x
x
F
w
l x
Linear Classifiers for Two Classes
• Fisher linear discriminant analysis
• Diagonal linear discriminant analysis (DLDA)
• Compound covariate predictor
• Golub’s weighted voting method
• Support vector machines with inner product kernel
• Perceptrons
Fisher LDA
( )
(1) (2)
,
log( ( ; , ) / ( ; , )) '
ix MVN
x x w x
The Compound Covariate Predictor (CCP)
• Motivated by J. Tukey, Controlled Clinical Trials, 1993
• A compound covariate is built from the basic covariates (log-ratios)
tj is the two-sample t-statistic for gene j.
xij is the log-expression measure of sample i for gene j.
Sum is over selected genes.
• Threshold of classification: midpoint of the CCP means for the two classes.
j
ijji xtCCP
Linear Classifiers for Two Classes
• Compound covariate predictor
Instead of for DLDA
(1) (2)
ˆi i
ii
x xw
(1) (2)
2ˆi i
ii
x xw
Support Vector Machine
2i
i
( )
j
minimize w
subject to ' 1
where y 1 for class 1 or 2.
jjy w x b
Perceptrons
• Perceptrons are neural networks with no hidden layer and linear transfer functions between input output– Number of input nodes equals number of genes selected– Number of output nodes equals number of classes minus 1– Number of inputs may be major principal components of genes
or major principal components of informative genes
• Perceptrons are linear classifiers
When p>>n
• It is always possible to find a set of features and a weight vector for which the classification error on the training set is zero.
• There is generally not sufficient information in p>>n training sets to effectively use more complex methods
Myth
• Complex classification algorithms such as neural networks perform better than simpler methods for class prediction.
• Comparative studies have indicated that simpler methods usually work as well or better for microarray problems because they avoid overfitting the data.
Other Simple Methods
• Nearest neighbor classification
• Nearest k-neighbors
• Nearest centroid classification
• Shrunken centroid classification
Nearest Neighbor Classifier
• To classify a sample in the validation set as being in outcome class 1 or outcome class 2, determine which sample in the training set it’s gene expression profile is most similar to.– Similarity measure used is based on genes
selected as being univariately differentially expressed between the classes
– Correlation similarity or Euclidean distance generally used
• Classify the sample as being in the same class as it’s nearest neighbor in the training set
Evaluating a Classifier
• Most statistical methods were not developed for p>>n prediction problems
• Fit of a model to the same data used to develop it is no evidence of prediction accuracy for independent data
• Demonstrating statistical significance of prognostic factors is not the same as demonstrating predictive accuracy
• Testing whether analysis of independent data results in selection of the same set of genes is not an appropriate test of predictive accuracy of a classifier
Internal Validation of a Classifier
• Re-substitution estimate– Develop classifier on dataset, test predictions
on same data– Very biased for p>>n
• Split-sample validation
• Cross-validation
Split-Sample Evaluation
• Training-set– Used to select features, select model type, determine
parameters and cut-off thresholds
• Test-set– Withheld until a single model is fully specified using
the training-set.– Fully specified model is applied to the expression
profiles in the test-set to predict class labels. – Number of errors is counted
Leave-one-out Cross Validation
• Omit sample 1– Develop multivariate classifier from scratch on
training set with sample 1 omitted– Predict class for sample 1 and record whether
prediction is correct
Leave-one-out Cross Validation
• Repeat analysis for training sets with each single sample omitted one at a time
• e = number of misclassifications determined by cross-validation
• Subdivide e for estimation of sensitivity and specificity
• With proper cross-validation, the model must be developed from scratch for each leave-one-out training set. This means that feature selection must be repeated for each leave-one-out training set.
– Simon R, Radmacher MD, Dobbin K, McShane LM. Pitfalls in the analysis of DNA microarray data. Journal of the National Cancer Institute 95:14-18, 2003.
• The cross-validated estimate of misclassification error is an estimate of the prediction error for model fit using specified algorithm to full dataset
Prediction on Simulated Null Data
Generation of Gene Expression Profiles
• 14 specimens (Pi is the expression profile for specimen i)
• Log-ratio measurements on 6000 genes
• Pi ~ MVN(0, I6000)
• Can we distinguish between the first 7 specimens (Class 1) and the last 7 (Class 2)?
Prediction Method
• Compound covariate prediction
• Compound covariate built from the log-ratios of the 10 most differentially expressed genes.
Number of misclassifications
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Pro
po
rtio
n o
f sim
ula
ted
da
ta s
ets
0.00
0.05
0.10
0.90
0.95
1.00
Cross-validation: none (resubstitution method)Cross-validation: after gene selectionCross-validation: prior to gene selection
Major Flaws Found in 40 Studies Published in 2004
• Inadequate control of multiple comparisons in gene finding– 9/23 studies had unclear or inadequate methods to deal with
false positives• 10,000 genes x .05 significance level = 500 false positives
• Misleading report of prediction accuracy– 12/28 reports based on incomplete cross-validation
• Misleading use of cluster analysis – 13/28 studies invalidly claimed that expression clusters based on
differentially expressed genes could help distinguish clinical outcomes
• 50% of studies contained one or more major flaws
Myth
• Split sample validation is superior to LOOCV or 10-fold CV for estimating prediction error
Comparison of Internal Validation MethodsMolinaro, Pfiffer & Simon
• For small sample sizes, LOOCV is much more accurate than split-sample validation– Split sample validation over-estimates prediction error
• For small sample sizes, LOOCV is preferable to 10-fold, 5-fold cross-validation or repeated k-fold versions
• For moderate sample sizes, 10-fold is preferable to LOOCV
• Some claims for bootstrap resampling for estimating prediction error are not valid for p>>n problems
Simulated Data40 cases, 10 genes selected from 5000
Method Estimate Std Deviation
True .078
Resubstitution .007 .016
LOOCV .092 .115
10-fold CV .118 .120
5-fold CV .161 .127
Split sample 1-1 .345 .185
Split sample 2-1 .205 .184
.632+ bootstrap .274 .084
Simulated Data40 cases
Method Estimate Std Deviation
True .078
10-fold .118 .120
Repeated 10-fold .116 .109
5-fold .161 .127
Repeated 5-fold .159 .114
Split 1-1 .345 .185
Repeated split 1-1 .371 .065
DLBCL Data
Method Bias Std Deviation MSE
LOOCV -.019 .072 .008
10-fold CV -.007 .063 .006
5-fold CV .004 .07 .007
Split 1-1 .037 .117 .018
Split 2-1 .001 .119 .017
.632+ bootstrap -.006 .049 .004
Permutation Distribution of Cross-validated Misclassification Rate of a
Multivariate Classifier• Randomly permute class labels and repeat
the entire cross-validation• Re-do for all (or 1000) random
permutations of class labels• Permutation p value is fraction of random
permutations that gave as few misclassifications as e in the real data
Gene-Expression Profiles in Hereditary Breast Cancer
• Breast tumors studied:7 BRCA1+ tumors8 BRCA2+ tumors7 sporadic tumors
• Log-ratios measurements of 3226 genes for each tumor after initial data filtering
cDNA MicroarraysParallel Gene Expression Analysis
RESEARCH QUESTIONCan we distinguish BRCA1+ from BRCA1– cancers and BRCA2+ from BRCA2– cancers based solely on their gene expression profiles?
BRCA1
g
# of
significant genes
# of misclassified
samples (m)
% of random permutations with
m or fewer misclassifications
10-2 182 3 0.4 10-3 53 2 1.0 10-4 9 1 0.2
BRCA2
g # of significant
genesm = # of misclassified elements
(misclassified samples)
% of randompermutations with m
or fewermisclassifications
10-2 212 4 (s11900, s14486, s14572, s14324) 0.810-3 49 3 (s11900, s14486, s14324) 2.210-4 11 4 (s11900, s14486, s14616, s14324) 6.6
Classification of BRCA2 Germline Mutations
Classification Method LOOCV Prediction Error
Compound Covariate Predictor 14%
Fisher LDA 36%
Diagonal LDA 14%
1-Nearest Neighbor 9%
3-Nearest Neighbor 23%
Support Vector Machine
(linear kernel)
18%
Classification Tree 45%
Does an Expression Profile Classifier Predict More Accurately Than Standard
Prognostic Variables?
• Some publications fit logistic model to standard covariates and the cross-validated predictions of expression profile classifiers
• This is valid only with split-sample analysis because the cross-validated predictions are not independent
log ( ) ( | )i iit p y x i z
Does an Expression Profile Classifier Predict More Accurately Than Standard
Prognostic Variables?
• Not an issue of which variables are significant after adjusting for which others or which are independent predictors– Predictive accuracy and inference are different
• The predictiveness of the expression profile classifier can be evaluated within levels of the classifier based on standard prognostic variables
Survival Risk Group Prediction
• LOOCV loop:– Create training set by omitting i’th case
• Develop PH model for training set
• Compute predictive index for i’th case using PH model developed for training set
• Compute percentile of predictive index for i’th case among predictive indices for cases in the training set
Survival Risk Group Prediction
• Plot Kaplan Meier survival curves for cases with predictive index percentiles above 50% and for cases with cross-validated risk percentiles below 50%– Or for however many risk groups and
thresholds is desired
• Compute log-rank statistic comparing the cross-validated Kaplan Meier curves
Survival Risk Group Prediction
• Evaluate individual genes by fitting single variable proportional hazards regression models to log expression for each gene
• Select genes based on p-value threshold for single gene PH regressions
• Compute first k principal components of the selected genes
• Fit PH regression model with the k pc’s as predictors. Let b1 , …, bk denote the estimated regression coefficients
• To predict for case with expression profile vector x, compute the k supervised pc’s y1 , …, yk and the predictive index = b1 y1 + … + bk yk
Survival Risk Group Prediction
• Repeat the entire procedure for permutations of survival times and censoring indicators to generate the null distribution of the log-rank statistic– The usual chi-square null distribution is not valid
because the cross-validated risk percentiles are correlated among cases
• Evaluate statistical significance of the association of survival and expression profiles by referring the log-rank statistic for the unpermuted data to the permutation null distribution
Sample Size Planning References
• K Dobbin, R Simon. Sample size determination in microarray experiments for class comparison and prognostic classification. Biostatistics 6:27-38, 2005
• K Dobbin, R Simon. Sample size planning for developing classifiers using high dimensional DNA microarray data. Biostatistics 2007
Sample Size Planning for Classifier Development
• The expected value (over training sets) of the probability of correct classification PCC(n) should be within of the maximum achievable PCC()
Probability Model
• Two classes• Log expression or log ratio MVN in each class with
common covariance matrix• m differentially expressed genes• p-m noise genes• Expression of differentially expressed genes are
independent of expression for noise genes• All differentially expressed genes have same inter-class
mean difference 2• Common variance for differentially expressed genes and
for noise genes
Classifier
• Feature selection based on univariate t-tests for differential expression at significance level
• Simple linear classifier with equal weights (except for sign) for all selected genes. Power for selecting each of the informative genes that are differentially expressed by mean difference 2 is 1-(n)
• For 2 classes of equal prevalence, let 1 denote the largest eigenvalue of the covariance matrix of informative genes. Then
1
( )m
PCC
1
1( ) 1
1
mmPCC n
m p m
Optimal significance level cutoffs for gene selection. 50 differentially expressed genes
out of 22,000 genes on the microarrays 2δ/σ n=10 n=30 n=50
1 0.167 0.003 0.00068
1.25 0.085 0.0011 0.00035
1.5 0.045 0.00063 0.00016
1.75 0.026 0.00036 0.00006
2 0.015 0.0002 0.00002
BRB-ArrayTools
• Contains analysis tools that I have selected as valid and useful
• Analysis wizard and multiple help screens for biomedical scientists
• Imports data from all platforms and major databases
• Automated import of data from NCBI Gene Express Omnibus
Predictive Classifiers in BRB-ArrayTools
• Classifiers– Diagonal linear discriminant– Compound covariate – Bayesian compound covariate– Support vector machine with
inner product kernel– K-nearest neighbor
– Nearest centroid– Shrunken centroid (PAM)– Random forrest– Tree of binary classifiers for k-
classes
• Survival risk-group– Supervised pc’s
• Feature selection options– Univariate t/F statistic– Hierarchical variance option– Restricted by fold effect– Univariate classification power– Recursive feature elimination– Top-scoring pairs
• Validation methods– Split-sample– LOOCV– Repeated k-fold CV– .632+ bootstrap
Selected Features of BRB-ArrayTools• Multivariate permutation tests for class comparison to control
number and proportion of false discoveries with specified confidence level– Permits blocking by another variable, pairing of data, averaging of
technical replicates• SAM
– Fortran implementation 7X faster than R versions• Extensive annotation for identified genes
– Internal annotation of NetAffx, Source, Gene Ontology, Pathway information
– Links to annotations in genomic databases• Find genes correlated with quantitative factor while controlling
number of proportion of false discoveries• Find genes correlated with censored survival while controlling
number or proportion of false discoveries• Analysis of variance
Selected Features of BRB-ArrayTools
• Gene set enrichment analysis. – Gene Ontology groups, signaling pathways, transcription
factor targets, micro-RNA putative targets– Automatic data download from Broad Institute– KS & LS test statistics for null hypothesis that gene set is not
enriched– Hotelling’s and Goeman’s Global test of null hypothesis that
no genes in set are differentially expressed– Goeman’s Global test for survival data
• Class prediction– Multiple classifiers– Complete LOOCV, k-fold CV, repeated k-fold, .632
bootstrap– permutation significance of cross-validated error rate
Selected Features of BRB-ArrayTools
• Survival risk-group prediction– Supervised principal components with and without clinical
covariates– Cross-validated Kaplan Meier Curves– Permutation test of cross-validated KM curves
• Clustering tools for class discovery with reproducibility statistics on clusters– Internal access to Eisen’s Cluster and Treeview
• Visualization tools including rotating 3D principal components plot exportable to Powerpoint with rotation controls
• Extensible via R plug-in feature• Tutorials and datasets
BRB-ArrayTools
• Extensive built-in gene annotation and linkage to gene annotation websites
• Publicly available for non-commercial use– http://linus.nci.nih.gov/brb
BRB-ArrayToolsMay 2007
• 7188 Registered users• 1962 Distinct institutions • 68 Countries• 365 Citations
• Registered users– 3951 in US
• 565 at NIH– 275 at NCI
• 2014 US EDU• 754 US Govt (non NIH)
– 3237 Foreign
Countries With Most BRB ArrayTools Registered Users
• France 270• Canada 269• UK 244• Germany 239• Italy 216• Taiwan 196• Netherlands 177• Korea 168• Japan 153• China 150
• Spain 146• Australia 130• India 107• Belgium 83• New Zeland 61• Sweden 50• Singapore 46• Brazil 48• Israel 41• Denmark 40
Acknowledgements
• Kevin Dobbin• Alain Dupuy• Wenyu Jiang• Annette Molinaro• Ruth Pfeiffer• Michael Radmacher• Joanna Shih• Yingdong Zhao• BRB-ArrayTools Development Team