Tools for core promoter prediction

1
Tools for core promoter prediction

• Thomas Abeel
• Department of Plant Systems Biology, VIB, Ghent
  University

2
Overview

• Introduction
• Core promoter prediction
• Physical properties of DNA
• Tools for promoter prediction
• EP3
• ProSOM
• Validation on the human genome

3
Introduction

• Core promoter
• Physical properties of DNA

4
Core promoter

• Promoter controlling element of a gene when,
  where, how much expression
• Core promoter binds transcription machinery,
  small (50 bp), contains transcription start site
  (TSS)
• Prediction identify these regions using
  computational resources

5

• Core promoter is the region 50 nt upstream of
  the TSS

6
Promoter prediction challenges

• A lot of unannotated sequence data is generated
• Experimental identification of promoters is
  expensive money, labor, time
• Computational techniques to identify promoters
  based only on sequence
• Help annotation projects to identify genes
• Help lab people to do targeted experiments

7
Physical properties of DNA

• Physical and chemical properties
• Examples bendability, denaturation energy
• Conversion tables
• Dinucleotide value
• Experimentally determined
• Replace each dinucleotide with value
• Average
• with a sliding window (in case of genome DNA)
• with multiple sequences

8
Example conversion table
9
Sample conversion

• AGTTA ? 85.12 108.8 66.51 50.11
• GGAAT ? 99.31 80.03 66.51 72.29
• CTGCA ? 85.12 64.92 135.83 64.92
• Profile ? 88.85 84.58 89.62 62.44

10
DNA denaturation energy

• Thermal stability of DNA
• The amount of energy that is required to separate
  two strands of DNA
• High values indicate stable regions
• Low values indicate unstable regions
• Dinucleotide conversion table
• Correlated with GC-content

11
Different properties (human)

• X-axis position relative to TSS
• Y-axis normalized average of the physical
  property
• 30,000 sequences

12
Small versus large scale (human)
13
Promoter elements
14
Different organisms (1)
15
Different organisms (2)
16
Introduction summary

• Core promoter has unique features
• Physical properties
• Large and small scale
• Valid for various organisms
• Valid irrespective of promoter elements
• Two unstable locations next to TSS
• Stable in large region

17
Promoter prediction tools

• EP3
• ProSOM

18
EP3
19
Sliding window conversion

• AGTTACTGCAGGAAT
• 4 nt sliding window
• ? 85.12 108.8 66.51 50.11 ...
• 86.81

20
On a genomic scale

• 400 nt sliding window
• 1 Mbp of chromosome 21
• Yellow ? forward strand genes
• Blue ? reverse strand genes
• Threshold on the profile for predictions

21
EP3 summary

• Large scale properties
• DNA denaturation
• 400 bp sliding window
• Threshold on the average
• value

22
ProSOM

• Small scale features
• DNA denaturation
• Inverse peak at -35
• Related to TBP
• Instability of 10 bp
• Inverse peak at 0
• Related to TSS

23
ProSOM

• Do not use profile directly
• Cluster known promoters and other sequences using
  an unsupervised method
• Clustering method separates promoters from other
  sequences
• New sequences that map to clusters with promoter
  content gt 75 are putative promoters

24
Training data

• DBTSS
• 30,000 promoter sequences -200,50
• Ensembl
• 30,000 transcribed sequences start5k,end-5k
• 30,000 intergenic sequences start5k,end-5k

25
Self-organizing maps

• Type of neural network
• Uses unsupervised learning competitive learning
• Maps high-dimensional input space to a lower
  dimension (2), while preserving topological
  properties
• Clustering based on the full physical profile of
  the sequences

26
Training procedure

• Randomly initialize 36 clusters (6x6 grid)
• Each sequence is added to the algorithm
• Features full physical profile of the sequence
• Compare this profile to the existing clusters
• Update the best matching cluster and its
  neighbors to look more like the added profile
• Identify clusters with high promoter content

27
Clustering result
28
Tools summary

• EP3
• Prediction based on large-scale physical
  properties
• Very simple ? fast, no training
• 400 bp sliding window
• ProSOM
• Predictions based on small-scale physical
  properties (indirectly)
• Slower, requires some training (unsupervised)
• 250 bp sliding window

29
Validation of EP3 and ProSOM

• Human genome
• Proof-of-concept on other genomes

30
Validation

• How many of our predictions correspond to known
  core promoters?
• For predictions that do not have a known
  associated core promoter, is there other
  evidence?
• How many of the known core promoters can we
  predict?
• Why are some known core promoters missed?

31
Human validation data

• Genome sequences for the complete human genome
• Coordinates of 123,400 known core promoters,
  transcription start sites (TSS)
• Determined in the lab
• Transcribed fragments and other evidence for
  ENCODE regions
• Predictions made by ProSOM and EP3

32
Quantifying right and wrong

• Counting scheme
• True positive (TP) correctly predicted TSS
• False positive (FP) false predictions, i.e.
  predictions not associated with a known TSS
• False negative (FN) unpredicted TSS, i.e.
  transcription start sites without prediction

33
Validation - graphical
TSS (FN)
TSS (TP)
-500 bp
500 bp
-500 bp
500 bp
FP
FP
Prediction
Region near TSS
Transcription Start Site
Region far from TSS
34
Performance measure

• Recall
• TP/(total tss)
• of sites recovered
• Precision
• TP/(total predictions)
• of predictions that is correct
• F-measure harmonic mean of recall and precision

35
Tool performance

• Among the best available
• Compared with gt10 state-of-the-art tools
• Preference for CpG-islands promoters, i.e.
  house-keeping genes

36
Performance ENCODE

• 85 of predictions have evidence for
  transcription within 50 bp
• Transcribed fragments
• CAGE tags
• ? Some false predictions in the genome are likely
  to be associated with transcription

37
Answers to the questions (1)

• How many of our predictions correspond to known
  core promoters?
• 66
• For predictions that do not have a known
  associated core promoter, is there other
  evidence?
• 85 within 50 bp, 98 within 500 bp

38
Answers to the questions (2)

• How many of the known core promoters can we
  predict?
• 34 (EP3) - 38 (ProSOM)
• Why are some known core promoters missed?
• Preference for GC-rich promoters
• Threshold is trade-off between recall and
  precision
• The profile is an average, not every promoter
  follows this pattern

39
Proof-of-concept on other organisms

• Whole genome sequences
• Gene annotation ? 5UTR end gene TSS
• EP3 12 genomes
• P. falciparum, O. pacifica, O. tauri, A.
  thaliana, O. sativa, P. trichocarpa, S.
  cerevisiae, S. pombe, C. elegans, D.
  melanogaster, T. nigroviridis, M. musculus
• ProSOM mouse (without retraining)

40
Why does it work?
41
ProSOM - EP3

• Human 48
• Mouse 45

42
Caveats

• 500 bp distance is large for plant and fungal
  genomes
• Prediction coverage is larger due to smaller
  genomes and higher gene density
• Protists genomes do not work because the profile
  is nearly flat
• Only gene annotation, we have no real core
  promoter data sets

43
Summary validation

• Two promoter predictors EP3 and ProSOM
• Works well on the human genome
• Recall 35, precision 66
• 98 of predictions have evidence
• Works on other eukaryotic genomes
• Some better than others
• Some caveats apply, only proof-of-concept

44
Conclusions

• Unique physical properties of core promoter
• EP3 (large scale)
• Additional validation on 12 eukaryotic genomes
• Requires no training, only threshold tuning
• ProSOM (small scale)
• Requires some training, but not species specific
• Extensively validated on the human genome
• Proof-of-concept on some other genomes

45
References and contact

• Abeel, T., Saeys, Y., Bonnet, E., Rouzé, P., Van
  de Peer, Y. (2008) Generic eukaryotic core
  promoter prediction using structural features of
  DNA. Genome Research 18, 310-23.
• Abeel, T., Saeys, Y., Rouzé, P., Van de Peer, Y.
  (2008) ProSOM Core promoter prediction based on
  unsupervised clustering of DNA physical profiles.
  Bioinformatics 24, i24-i31.
• E-mail thomas.abeel_at_psb.ugent.be
• Web http//bioinformatics.psb.ugent.be/