Data Mining the Yeast Genome Expression and Sequence Data Alvis Brazma European Bioinformatics Institute.

Data Mining the Yeast Genome Expression and

Sequence Data

Alvis Brazma

European Bioinformatics Institute

Why the yeast is interesting to the industry

Easy to work with (first) fully sequenced eukaryotic model organism

30% of genes have analogs in human most known human disease genes

have homologues in the yeast for food industry interesting in itself

Genetic networks

promoter1 gene1 promoter2 gene2 promoter3 gene3 promoter4 gene4DNA

RNA

transcription

translation

proteins

transcription factors

Mining the Yeast Expression Data

The long term goals:» reconstructing the gene regulation networks and

relating it to metabolic pathways Short term goals:

» correlating gene expression profiles with gene functional classes and using this for prediction of gene functions

» correlating gene expression profiles with promoter regions

Yeast microarray

Yeast gene expression during diauxic shift (DeRisi et al)

Yeast cells from an exponentially growing yeast culture were inoculated into fresh medium and after some initial period were harvested at seven 2-hour intervals. Their mRNA were isolated, and fluorescently labeled cDNA prepared. Two different fluorescents were used - one from cells harvested in each of the successive time-points, other from the cells harvested at the first time-point (reference measurement). The cDNA from each time-point together with the reference cDNA were hybridized to the microarray with approximately 6400 DNA sequences representing ORFs of the yeast genome. Measurements of the relative fluerescence intensity for each element reflect the relative abundance of the corresponding mRNA.

Visualizing the data (expression profile of the

“first” 250 genes)

Average expression level of genes at the respective time-

points

Three approaches

Finding correlations between gene expression profiles and their functional classes

Building decision trees for predicting gene functional classes from their expression data

In silico discovery of putative transcription factor binding sites in the regions upstream to the genes with similar expression profiles (to appear in Genome Research, Dec. 1998)

Gene distribution across the functional classes

Energy gene subclasses in the yeast (less frequent merged in

one)

Gene expression for energy genes during the diauxic shift at the seven time-points

Expression profiles of respiration genes

Expression profiles of fermentation genes

Average expression levels at the 7 time-points and for energy class genes during

diauxic shift

Average expression levels at all time-points and for all energy

classes

Energy classes distribution

Energy classes distribution

Decision tree for respiration genes

Decision tree for fermentation

Tricarboxilacid, respiration and reserves decision tree

Clustering the gene expression profiles by

discretization of gene expression measurment space

Logarithm ofexpression ratio

Time points

1 2 3 4 5 6 7

Corresponding discrete pattern: 000012-1Put the genes mapping to the same discrete pattern in a cluster

0

1

2

-1

-2

Organization of a typical yeast promoter

URS URS TATA I

Coding Region

40 - 120 bp

20 - 700 bp

RNA

40 - 60 bp

In silico discovery of transcription factor binding sites from expression data

Take data from gene expression level measurements (from DNA array technologies) ->

Cluster together genes with similar expression profiles ->

Take sequences upstream from the genes in each cluster ->

Look for sequence patterns overrepresented in a cluster

Clustering genes by similar expression profiles

Put in each cluster all genes that map to the same discrete pattern

Different thresholds give different clustering systems

We obtained 32 different clusters containing from 10 to 77 genes and 11 clusters containing at least 25 genes

Hypothesis to test

Genes with similar expression profiles may be regulated by similar expression mechanisms and thus may contain similar transcription factor binding sites

Discovering regulatory elements in gene upstream

sequences

Take the sequences of a certain length (e.g., 300 bp) upstream to all genes with a certain expression profile

Look for a priori unknown sequence patterns that are over-represented in these regions (taking into account the other upstream regions as background)

Pattern discovery in bioseqeucnes

Group together sequences thought to have common biological (structural, functional) properties, ignoring the purely sequence (syntactic) properties

Study the purely syntactic properties of these sequences ignoring their biological (semantic) properties.

Problem of “noise”

Gene expression measurement accuracy is bout factor of 2 (in 95% cases)

Clusters very dependant on the clustering method or thresholds

The same expression profile does not necessarily mean the same regulation mechanism

Dealing with noise

One cannot look for patterns common to the set of strings, but for patterns overrepresented in the set

looking for sets of patterns covering the set

Use of “negative” or background setquences

More powerful algorithms than the currently existing

are needed

We used such new, more powerful algorithm, based on suffix-tree representation of the sequence space (implemented by Jaak Vilo at Helsinki University)

We looked systematically for all patterns discriminating the upstream regions in the clusters from randomly selected upstream regions

Use of negative sequences

Looking for patterns that are overrepresented in the sequences upstream from genes in a cluster in comparison to all other upstream sequences

The rating function

Given two sets S+ and S- and a pattern P, return rating R(S+, S-,P)

Two rating functions that we used:» ratio: nr of sequences in S+ matching P divided

by nr of sequences in S- matching P

» probability that the pattern can occur in S+ “by chance” assuming that the occurrences in S- are “by chance” and using binomial distribution

The sequence pattern discovery experiment

We run the algorithm on upstream sequences (length 2 * 300) of all the 32 gene clusters

Each cluster produced hundreds of overrepresented patterns

The problem of validation

Some discovered sequence patterns from clusters of upstream sequences

Clusters with the increase in the expression level after time-point 6:

CCCCT - known to be a stress responsive motif

Clusters with the decrease in the expression level after time-point 6:

ATCC..T..A - RAP1 protein

ATC..TAC - RAP1, REB1, BAF1

ATTTCA…T - GA-BF protein

Statistical validation of the discovered patterns

For each cluster choose a random set of upstream regions of the same number

Run the pattern discovery algorithm on the random regions set in addition to the cluster

Compare the scores of the discovered patterns from the cluster and random set

Conclusions

The discovered patterns are in accordance with the existing knowledge

Transcription factor binding sites can be discovered in silico from gene expression data

More refined and validated gene expression measurements are needed

Acknowledgements

Inge Jonassen (Bergen) Jaak Vilo , Esko Ukkonen (Helsinki) Alistair Ewing, Neil Skilling (Quadstone

Ltd - developers of Decisionhouse data mining software)

BIOVIS and BIOSTANDARDS projects from the EU at EBI