Bioinformatics/Computational Biological Applications of ... - NICTA


Oct 2, 2013 (3 years and 8 months ago)


Bioinformatics Applications of
Machine Learning

Brian Parker

NICTA Life Sciences


Bioinformatics/computational biology: data analysis of
molecular biology datasets

Aims of this lecture: To introduce

Some background molecular biology and biotechnology

e.g. microarrays, expressed sequence tags (EST’s)

Some bioinformatics applications of the machine
learning methods covered in the lectures so far, and
some of the issues and caveats specific to such

Overview cont’


Unsupervised and supervised
classification of expression microarrays

Clustering of EST data and EST sequence
alignment and discussion of genomic
distance measures

Background molecular biology

Central dogma of molecular biology:

> transcribed
> translated

Protein has certain tertiary structures to
carry out function e.g. structural elements,
enzymes for metabolic processes, gene
regulation etc.

Background molecular biology cont’

DNA is double
stranded polymer of 4 nucleotides
Adenine(A), Cytosine (C), Guanine (G), Thymine (T)

A gene is a segment of DNA coding for a protein.

mRNA is single

Protein is polymer of 20 amino acids

The genetic code maps from the 4
letter alphabet of
DNA to the 20

letter alphabet of protein

Note: Recent extension of central dogma


not translated into protein and directly regulate
expression of other genes

Background molecular biology cont’

These stages lead to several higher

Gene regulatory networks, pathways

protein interaction networks

Biochemical networks


throughput data analysis

“Omics” = high throughput datasets

Following the central dogma, we have:


from high
throughput sequencing of DNA


from high
throughput sequencing of
RNA and transcribed genome


from high
throughput analysis of protein


from analysis of biochemical metabolites

Microarray technology

Simultaneously measure the expression of
10s of thousands of genes.

Several technologies e.g. Spotted and
oligonucleotide arrays (Affymetrix)

Large array of probes designed as a
complementary match to the transcript of

Microarray technology

Relies on hybridization

i.e. single
nucleic acids bind to their complement.

mRNA extracted
> reverse transcriptase
cDNA (biotin

> hybridize to array
> scan image (amount of
fluorescence relates to amount of mRNA)

> convert to expression levels.

Important to normalize arrays to remove
variations due to differing lab technique (not
covered in this lecture).

Spotted array image

Affymetrix array


large p, small n dataset, where n is the
number of samples and p is the number of
features e.g. 50,000 genes, 100 patient
samples is typical

This is the opposite assumption of earlier
statistical and machine learning


Can lead to novel problems:

(1) Many techniques assume n <= p e.g. LDA
cannot be applied directly as covariance matrix
is under
determined and can not be estimated,
so feature selection is required.

(Even where a method e.g. SVMs can handle the
high dimensionality, feature selection is still
useful to remove noise genes).


(2) Large opportunity for selection bias to occur in
feature selection.

(3) Large multiple hypothesis correction problem.
How to do this without being too conservative?

(Note: we will be talking about expression
arrays; there are other array types such as SNP
arrays that hybridize with genomic DNA to
measure copy number, LOH etc)

Microarray Analysis

3 broad problems in microarray analysis
(Richard Simon):

class discovery (unsupervised classification)

(2) class comparison (differential gene expression)

(3) class prediction (supervised classification)

Hierarchical clustering

heat map

E.g. Sorlie et. al. (2001) reported several previously
unidentified subtypes of breast cancer using clustering.

(Sorlie et al, “Gene expression patterns of breast
carcinomas distinguish tumor subclasses with clinical
implications”, PNAS)

Filter methods

Specific versus non
specific filtering

specific filtering doesn’t use the class
labels but removes noise genes of low
variance etc.

N.B. in clustering, don’t do specific filtering
and then cluster!

Specific Filtering

statistic (one
statistic if > 2 samples)

problem is that there often isn’t enough data to estimate

Fold change: simplest method; ratio of expression levels

(but as microarray data is typically log transformed,
calculated as difference of means)

Specific Filtering cont’

Moderated t
statistic. Estimate variance across multiple

Many different versions of moderated variations on the t
test (e.g. regularized t
test of Smyth (2004) (Limma
package in Bioconductor), SAM).

They combine a gene
specific variance estimate with an
overall predicted variance (e.g. the microarray average)
i.e. roughly

Where is some measure of group difference (e.g. difference of means)

is a predicted variance based on all genes, (may be transformed) and

is estimated variance based on the particular gene.

B is a “shrinkage factor” that ranges from 0 to 1.

For B = 1, denominator is effectively constant and so we get the fold change.

For B = 0, standard t
test without any shrinkage.

in experiment results

Experiment with very small spike
in set (6

(ref. Bioinformatics and Computational Biology Solutions Using R
and Bioconductor)

t better than fold
change better
than t

Embedded and wrapper methods

Wrapper method uses an outer cross

select gene set with smallest loss.

Full combinatorial search is too slow

need to
do forward or backward feature selection

Embedded e.g. Recursive feature elimination
(RFE) (Guyon and Vapnik). Uses SVM internal
weights to rank features

removes worst feature
and then iterate. (original paper had a severe
selection bias).

Differential gene expression

multiple hypothesis testing

Setting a limit with p
value = 0.05 is too lax due to
multiple hypothesis testing.

Doing a multiple hypothesis correction such as
Bonferroni correction (multiply p
value by number of
genes) is too conservative. In practice, some in
value may be chosen empirically.

This is controlling family
wise error rate (FWER)

the p
value threshold so whole study has a defined false
positive rate. For an exploratory study such as
differential gene expression, we are willing to accept a
higher false positive rate.

False Discovery Rate (FDR)

In this case, what we really want is to specify the
proportion of false positives we will accept
amongst the gene set we have selected as

the false discovery rate FDR.

Several variants of FDR

an example is the q
value of Storey and Tibshirani.

F = false positives, T = true positives, S = “significant features”

Class Prediction

Can be a classification problem e.g. cancer vs
normal or a regression problem, e.g. survival

Simple methods work well in practice due to
small patient numbers.

Dudoit, Fridlyand and Speed compared K
various linear discriminants and CART.

Conclusion: k
nn and DLDA performed best, and
ignoring correlation between genes helped:
DLDA vs correlated LDA.

Selection bias in microarray studies

Because of the high dimensionality and small sample size of
microarray data, it is very likely that a random gene will by luck
correlate with the class labels.

So selecting the best gene set for classification will give an
optimistic bias if done outside of the cross
validation loop.

It is essential that when using cross
validation, the test set is not
used in any way in each fold of the cross validation. This means that
all feature selection and (hyper) parameter selection and model

be repeated for each fold.

Selection Bias cont’

(From Amboise and McLauclan “
Selection bias in gene extraction on the

basis of microarray gene
expression data”)

Gene set enrichment analysis

Previous approaches discussed were univariate
filter methods, essentially treating each gene

Looking at the overall difference in expression of
sets of genes that are known, by other
experiments, to be related ,e.g. part of the same
pathway or similar gene ontology (GO)
annotation, can be a more powerful test to find
significant differences.


Genes are ranked using a univariate metric

An enrichment score for the gene set is

using a Kologorov

The significance level of the enrichment score
is computed using a permutation test (where
the shuffled labels keep the gene set together).

A FDR is computed to correct for multiple
hypothesis testing.

EST analysis

Expressed sequence tags (ESTs) are
short, unedited, randomly selected single
pass sequence reads derived from cDNA
libraries. Low cost, high throughput.

(cDNA is generated by reverse
transcriptase applied to RNA)

EST analysis steps

(1) They need to be clustered into longer
consensus sequences (unsupervised

(2) They can then be sequence aligned
against the genome for gene
finding etc.

These two methods require different
genomic sequence distance measures…

Similarity measures for genomic

Most data analysis methods use some
underlying measure of similarity or distance
between samples either explicity or implicitly and
this is a major determinant of their performance

e.g. the hierarchical clustering discussed in
previous lectures typically has a (dis)similarity
matrix passed into the function so that the
particular similarity measure used is decoupled
from the clustering algorithm

Similarity measures for genomic

This idea can be generalized to supervised
classification and other data analysis

even when the similarity measure is
implicit, it can often be algebraically
manipulated to make it explicit

(and in this case is the measure is typically a
dot product

generalized by kernel
methods to be discussed in later lectures)

Similarity measures for genomic

So, it is important to generate good similarity
measures between genomic sequences.

Two broad classes:

Alignment methods and

free methods

Alignment methods

Model insertions/deletions and substitutions

form of edit distance


global alignment

Based on dynamic programming


local alignment (includes only
matching high
scoring regions)

BLAST uses a non
based heuristic to
quickly rule out bad matches

Used for sequence alignment and database

free methods

based distance measures
assume conservation of contiguity
between homologous segments

Not always the case e.g. ESTs from
different splice variants or genome

free methods

Based on comparing word frequencies


statistic = number of k
word matches
between two sequences.

Can be shown to be an inner product of
count vectors.

Useful for EST clustering

Other areas of bioinformatics

Several other areas of bioinformatics not
covered here which also use machine
learning techniques

Protein secondary and tertiary structure
and motif finding

De novo gene prediction by matching
known promoter and coding sequence