Biology, Computational Part 11 Genefinding
Robert F. Murphy Copyright 1997, 2001, 2003-2007. All rights reserved.
Clues to locations of genes (Prokaryotic Signals)
s
for Transcription
x Promoters x Transcription factor binding sites
s
for Translation
x Ribosome binding sites x Start/stop codons
Prokaryotic vs. Eukaryotic Genefinding
s
Simple to create programs to look for "grammatical" combination of prokaryotic signals s Much more complicated for eukaryotes due to the presence of introns and additional regulatory elements
Clues to locations of genes (Eukarytoic Signals)
s
for Transcription
x x x x x
Promoters Transcription terminators Topoisomerase II binding sites Topoisomerase I cleavage sites Transcription factor binding sites Donor and acceptor sites Branch points
s
for Splicing
x x
Clues to locations of genes (Eukaryotic Signals)
s
for mRNA Processing
x Polyadenylation sites
s
for Translation
x Ribosome binding sites x Start/stop codons
DNA->RNA->protein
Signal Sensors: Consensus Sequences
s
Simple
x TATA
s
PROSITE expression
x Y-x-G-A-[FL]-[KRHNQ]-C-L-x(3,4)-G-
[DENQ]-V-[GA]-[FYW] x (iron binding site in transferrin)
Signal Sensors: Networks
s
Profile, PSSM (equivalent to perceptron) s Neural Network (multi-layer)
Content Sensors: Coding Regions
s
GeneMark: 3 fifth-order Markov models
x one for each reading frame
s
GRAIL: uses neural net with inputs from
x coding potential measures x base composition x signal sensor output for flanking splice sites
Integrated Systems
s
Use dynamic programming to find best combination of signal/content sensors s Apply "linguistic" rules to say what parts are required and in what order
Gene model
B = gene start S = translation start D = donor A = accceptor T = translation stop E = gene end
Basic Implementation
s
Use an HMM to model what state Q each nucleotide from X is in (given parameters ) s Train HMM with known genes to estimate s For unknown sequence, find Q to maximize P(Q | X, s Used by GENSCAN, HMMgene
GENSCAN HMM
s
Handles genes on both forward and reverse strands
Adding Homology
s
Can try to include information databases from of known proteins to help decide whether an exon is coding s For each candidate exon, increase the score if there is homology with a known protein s This approach used by Genie, GeneID+, GeneParser3, Grail
Adding ESTs
s
Can try to include information from EST databases s EST (Expressed Sequence Tag) databases show sequences that are "known" to be present in mRNA (cDNA) s For each candidate exon, increase the score if it matches to an EST s Used by AAT, Grail
Drawbacks
s
Using homology or ESTs may bias results toward genes similar to known genes (homology) or highly expressed genes (ESTs)
Assessing Performance
s
1995: Burset and Guigo used benchmark set of 575 vertebrate genes to compare s 2001: Rogic et al. redid comparison with better test set of 195 genes (Genome Research 11:817-832, 2001)
x http://www.cs.ubc.ca/~rogic/evaluation/
Machine Learning 101
s s
Two types of learning methods Supervised
x
x
have examples of things in different "classes" that you want program to recognize or "classify" goal is for program to learn "rules" or "boundaries" to distinguish classes have program figure out what classes exist
s
Unsupervised
x
Machine Learning 101
s s s s
Assume that have a group of...