Transcript Discrimination Methods

Discrimination Methods
As Used In Gene Array Analysis
Discrimination Methods



Microarray Background
Clustering and Classifiers
Discrimination Methods:







Nearest Neighbor
Classification Trees
Maximum Likelihood Discrimination
Fisher Linear Discrimination
Aggregating Classifiers
Results
Conclusions
Microarray Background



Nowadays, very little is known about
genes functionality
Biologists provides experimental
information for analyze, in order to find
biological function to genes
Their tool - Microarray
Microarray Background

The process:




DNA samples are taken from the test subjects
Samples are dyed with fluorescent colors, and placed on
the Microarray, which is an array of DNA built for each
experiment
Hybridization of DNA and cDNA
The result:

Spots in the array are
dyed in shades of Red
to Green, relative to
their expression level
on the particular
experiment
Microarray Background

Sample 1
Sample 2
Gene 1
1.04
2.08
Gene 2
3.2
10.5
Gene 3
3.34
1.05
Gene 4
1.85
0.09
Microarray data is translated into an nxp table, where
p is the number of genes in the experiment, and n is
the number of samples
Clustering



What to do with all
this data?
Find clusters in the
nxp space
Easy in low
dimensions, but in
our multidimensional space,
it is much harder
example for clusters in 3D
Clustering
Why Clustering?



Find patterns in our
experiments
Connect specific
genes with specific
results
Mapping genes
Classifiers



The tool – Classifiers
Classifier is a function that splits the space into K
disjoint sets
Two approaches:

Supervised Learning (Discrimination Analysis):




Unsupervised Learning (Cluster Analysis):




K is known
learning set is used to classify new samples
used to classify malignancies into known classes
K is unknown
the data “organizes itself”
used for identification of new tumors
Feature Selection – another use for classifiers

used for identification of marker genes
Classifiers


We will discuss only about supervised
learning
Discrimination methods:





Fisher Linear Discrimination
Maximum Likelihood Discrimination
K Nearest Neighbor
Classification Trees
Aggregating classifiers
Nearest Neighbor




We use a predefined learning set, already
classified
New samples are being classified into the
same classes of the learning set
Each sample is classified its K nearest
neighbors, according to a distance metric
(usually Euclidian distance)
The classification is made by majority of
votes
Nearest Neighbor

NN, example
Nearest Neighbor
Cross-Validation:
 Method for finding the best K to use
 Test each of {1,...,T} as K, by running
the algorithm T times on a known test
set, and choosing the K which gives
the best results
Classification Trees



Partitioning of the space into K classes
Intuitively presented as a tree
Two aspects:



Constructing the tree from the training set
Using the tree to classify new samples
Two building approaches:


Top-Down
Bottom-Up
Classification Trees

Bottom-Up approach:


Start with n clusters
In each iteration:



merge the two closest clusters,
using a measure on clusters
Stop when a certain criteria is met
Measures on clusters:



minimum pairwise distance
average pairwise distance
maximum pairwise distance
Classification Trees
Bottom-Up approach, example
c3
c5
c1
c4
c2
c6
Classification Trees

Top-Down approach:

In each iteration:



Choose one attribute
Divide the samples space according to this
attribute
Use each of the sub-groups just created as
the samples space for the next iteration
Classification Trees
Top-Down approach, example
c3
c5
c1
c4
c2
c6
Classification Trees

Three main aspects of tree construction:

split selection rule
which attribute we should choose for splitting in each
iteration?

split stopping rule
when should we stop clustering?

class assignment rule
which class will each leaf represent?

Many variants:



CART (classification and regression trees)
ID3 (iterative dichotomizer)
C4.5 (Quinlan)
Classification Trees - CART

Structure


Binary tree
Splitting criterion

Gini index:



for a node t and classes (1,...,k),
2
let Gini index be GINI (t )  1   j P  j | t 
where P(j|t) is the relative part of class j at node t
Split by a minimized Gini index of a node
Stopping criterion

Relatively balanced tree
Classification Trees
Classify new samples, example
red
Left color
Right color
c1
c2
Right color
c3
c4
Right color
c5
c6
Classification Trees
Over Fitting:
 Bias-Variance trade-off



The deeper the
tree the bigger its
variance
The shorter the
tree the bigger
the bias
Balance trees will
give the best
results
Maximum Likelihood




Probabilistic approach
Suppose a training set is given, and we
want to classify a sample x
Lets compute the probability of a class ‘a’
when x is given, denoted as P(a|x).
Compute it for each of the K classes, and
assess x to the class with the highest
resulting probability:
C  x   argmax P  a | x 
a
Maximum Likelihood



Obstacle: P(a|x) is unknown
Solution: Bayes rule P a | x   P  x | a  P a 
P x
Usage:
C  x   argmax
a



P  x | a  P a 
P x
P(a) is fixed (the relative part of a in the test set)
P(x) is class independent so also fixed
P(x|a) is what we need to compute now
Maximum Likelihood



Remember that x is a sample of p genes:
x   x1,..., x p 
If the genes’ densities were independent, then
as a multiplication of the relative parts of samples
on each gene P ( x | a)  P  x1 | a   ...  P  x p | a 
Independence hypothesis:



makes computation possible
yields optimal classifiers when satisfied
but seldom satisfied in practice, as attributes (variables)
are often correlated
Maximum Likelihood



If the conditional densities of the classes are fully
known, a learning set is not needed
If the conditional densities are known, we still have
to find their parameters
More information may lead to some familiar results:

Densities with multivariate class densities

C  x   argmin ( x  k )k1( x  k )t  log k
k


Densities with diagonal covariance matrices
 ( x j   kj )2

2 
C  x   argmin  

log

kj 
2

k
j 1 

kj

p

Densities with the same diagonal covariance matrix
p
( x j  kj )2
j 1
 2j
C  x   argmin 
k
Fisher Linear Discrimination

Lower the problem from multidimensional to single-dimensional




Let ‘v’ be a vector in our space
Project the data on the vector ‘v’
Estimate the ‘scatterness’ of the data as
projected on ‘v’
Use this ‘v’ to create a classifier
Fisher Linear Discrimination


Suppose we are in a 2D space
Which of the three vectors is an optimal ‘v’?
Fisher Linear Discrimination

The optimal vector maximizes the ratio of
between-group-sum-of-squares to withingroup-sum-of-squares, denoted v Bv
t
v tWv
between
within
within
Fisher Linear Discrimination
Suppose a case two classes

Mean of these classes samples:
mi 
1
x
n xXi

Mean of the projected samples:
mi 
1
1
y

w t x  w t mi


n yYi
n xXi

‘Scatterness’ of the projected samples:

Criterion function:
J v  
m1  m2
s12  s 22
2
si 2 
 (y  m )
2
y Yi
i
Fisher Linear Discrimination


Criterion function should be maximized
Present J as a function of a vector ‘v’
Wi 
 ( x  m )( x  m )
t
i
x X i
i
W  W1  W2
si2 
 (v x  v m )
t
x X i
t
2
i

 v ( x  m )( x  m ) v  v W v
t
x X i
t
i
i
t
i
s12  s22  v tWv
B  (m1  m2 )(m1  m2 )t
(m1  m2 )2  (v t m1  v t m2 )2  v t (m1  m2 )(m1  m2 )t v  v t Bv
v t Bv
J v   t
v Wv
Fisher Linear Discrimination


The matrix version of the criterion
works the same for more than two
classes
J(v) is maximized when Bv  Wv
Fisher Linear Discrimination
Classification of a new observation ‘x’:
 Let the class of ‘x’ be the class whose
mean vector is closest to ‘x’ in terms of
the discriminant variables
 In other words, the class whose mean
vector’s projection on ‘v’ is the closest
to the projection of ‘x’ on ‘v’
Fisher Linear Discrimination
Gene selection



most of the genes in the experiment will not be
significant
reducing the number of genes reduces the error
rate, and makes computations easier
For example, selection by the ratio of each gene’s
between-groups and within-groups sum of squares
  I ( y  k )( x  x )
R
j



For each gene j, let
  I ( y  k )( x  x )
and select the genes with the larger ratio
2

i
k
i
kj
j
i
k
i
ij
kj
2
Fisher Linear Discrimination
Error reduction




Small number of samples makes the error more
significant
Noise will affect measurements of small values, and
thus the WSS can be too big in some
measurements
This will make the selecting criterion of a gene
bigger than its real importance to the discrimination
Solution - Adding a minimal value to the WSS
Aggregating Classifiers



A concept for enhancing performance
of classification procedures
A classification procedure uses some
prior knowledge (i.e. training set) to get
its classifier parameters
Lets aggregate these parameters from
more training sets into a stronger
classifier
Aggregating Classifiers

Bagging (Bootstrap Aggregating) algorithm



Generate B training sets from the original
training set, by replacing some of the data in
the training set with other data
Generate B classifiers, C1,...,Cb
Let x be a new sample to be classified.
The class of x is the majority class of x on the B
classifiers C1,...,Cb
Aggregating Classifiers

Boosting, example
T1
Classifier 1
T2
Classifier 2
Tb
Classifier b
Aggregated
classifier
training set
Aggregating Classifiers

Weighted Bagging algorithm





Generate B training sets from the original
training set, by replacing some of the data in
the training set with other data
Save the replaced data from each set as a
training set, T(1),...,T(b)
Generate B classifiers, C(1),...,C(b)
Give each classifier C(i) a weight w(i) according
to its accuracy on the test set T(i)
Let x be a new sample to be classified.
The class of x is the majority class of x on the B
classifiers C(1),...,C(b), with respect to the
weights w(1),...,w(b).
Aggregating Classifiers

Improved Boosting, example
T1
Classifier 1
T2
Classifier 2
training set
Weight
function
Tb
Classifier b
Aggregated
classifier
Imputation of Missing Data



Most of the classifiers need
information about each spot in the
array in order to work properly
Many methods of missing data
imputation
For example - Nearest Neighbor:

each missing value gets the majority
value of its K nearest neighbors
Results
Dudoit, Fridlyand and Speed (2002)

Methods tested:





Fisher Linear Discrimination
Nearest Neighbor
CART classification tree
Aggregating classifiers
Data sets:



Leukemia – Golub et al. (1999)
72 samples, 3,571 genes, 3 classes (B-cell ALL, T-cell ALL, AML)
Lymphoma – Alizadeh et al. (2000)
81 samples, 4,682 genes, 3 classes (B-CLL, FL, DLBCL)
NCI 60 – Ross et al. (2000)
64 samples, 5,244 genes, 8 classes
Results - Leukemia data set
Results - Lymphoma data set
Results - NCI 60 data set
Conclusions

“Diagonal” LDA: ignoring correlation
between genes improved error rates

Unlike classification trees and nearest
neighbors, LDA is unable to take into
account gene interactions

Although nearest neighbor is s simple and
intuitive classifier, its main limitation is that it
give very little insight into mechanisms
underlying the class distinctions
Conclusions

Classification trees are capable of handling
and revealing interactions between
variables

Variable selection: a crude criterion such as
BSS/WSS may not identify the genes that
discriminate between all the classes and
may not reveal interactions between genes

With larger training sets, expect
improvement in performance of aggregated
classifiers