4.2 Ab initio gene prediction

Ab initio gene prediction is a crucial step in genome annotation. It uses statistical models to identify genes in genomic sequences without relying on external evidence. These methods analyze DNA signals, sequence composition, and gene structure to predict potential coding regions.

Markov models, particularly hidden Markov models (HMMs) and generalized HMMs, form the backbone of ab initio prediction. These models are trained on known genes to learn patterns and probabilities. Tools like GENSCAN and Glimmer apply these models to predict genes in eukaryotic and prokaryotic genomes.

Fundamentals of ab initio gene prediction

• Ab initio gene prediction aims to identify genes in genomic sequences without relying on external evidence such as cDNA or protein sequences
• It is a crucial step in genome annotation and understanding the genetic basis of organisms
• The methods rely on statistical models trained on known gene structures to predict potential coding regions and splice sites

Biological basis for gene prediction

Signals in DNA sequence

Top images from around the web for Signals in DNA sequence

• 

  RNA splicing - Wikipedia View original

  Is this image relevant?

• 

  Gene Regulation in Prokaryotes | Biology for Majors I View original

  Is this image relevant?

• 

  Eukaryotic Transcription Gene Regulation | Biology for Non-Majors I View original

  Is this image relevant?

• 

  RNA splicing - Wikipedia View original

  Is this image relevant?

• 

  Gene Regulation in Prokaryotes | Biology for Majors I View original

  Is this image relevant?

1 of 3

Top images from around the web for Signals in DNA sequence

• 

  RNA splicing - Wikipedia View original

  Is this image relevant?

• 

  Gene Regulation in Prokaryotes | Biology for Majors I View original

  Is this image relevant?

• 

  Eukaryotic Transcription Gene Regulation | Biology for Non-Majors I View original

  Is this image relevant?

• 

  RNA splicing - Wikipedia View original

  Is this image relevant?

• 

  Gene Regulation in Prokaryotes | Biology for Majors I View original

  Is this image relevant?

1 of 3

• Promoter regions upstream of genes contain binding sites for transcription factors (TATA box, CAAT box)
• Translation start and stop codons (ATG, TAA, TAG, TGA) mark the beginning and end of coding regions
• Splice donor and acceptor sites (GT-AG) flank introns and are recognized by the spliceosome during pre-mRNA processing
• Polyadenylation signals (AATAAA) indicate the 3' end of transcripts and guide cleavage and polyadenylation

Sequence composition of genes

• Coding regions exhibit biased nucleotide composition compared to non-coding regions
• Codon usage bias reflects the preferential use of certain codons for amino acids due to tRNA abundance or translational efficiency
• CpG islands, regions with high CG content, are often associated with promoters and transcription start sites
• Repetitive elements (SINEs, LINEs) are less frequent in coding regions compared to intergenic regions

Markov models for gene prediction

Markov chains vs hidden Markov models

• Markov chains model the probability of a sequence of states based on the current state (nucleotide or codon)
• Hidden Markov models (HMMs) introduce hidden states (exon, intron, intergenic) that emit observable sequences with different probabilities
• HMMs allow for modeling the dependencies between adjacent states and the observed sequence

Training HMMs on known genes

• HMM parameters (transition and emission probabilities) are estimated from a training set of annotated genes
• The Baum-Welch algorithm is used for unsupervised training, iteratively updating parameters to maximize the likelihood of the observed sequences
• Supervised training with labeled data (exon, intron, intergenic) can improve the accuracy of the model

Viterbi algorithm for optimal path

• The Viterbi algorithm finds the most probable sequence of hidden states given an observed sequence and trained HMM
• It uses dynamic programming to efficiently compute the maximum likelihood path through the state space
• The optimal path corresponds to the predicted gene structure, with transitions between exon, intron, and intergenic states

Generalized hidden Markov models

GHMMs vs HMMs

• Generalized hidden Markov models (GHMMs) extend HMMs by allowing states to emit variable-length sequences
• In GHMMs, each state can have a duration distribution that models the length of the emitted sequence
• GHMMs are more suitable for modeling biological features with variable lengths, such as exons and introns

Duration modeling in GHMMs

• Duration distributions (geometric, gamma, or explicit) capture the length variability of features like exons and introns
• Incorporating duration modeling improves the accuracy of gene structure prediction by favoring biologically plausible lengths
• The duration distribution parameters are estimated from the training data along with transition and emission probabilities

Gene structure modeling with GHMMs

• GHMMs can model the complex structure of eukaryotic genes with multiple exons and introns
• States represent different gene components (promoter, 5' UTR, exon, intron, 3' UTR, polyadenylation site)
• Transitions between states capture the order and dependencies of gene components (exon-intron boundaries, splice sites)
• The GHMM architecture is designed to reflect the biological constraints and patterns of gene structure

Ab initio gene prediction tools

GENSCAN for eukaryotic gene prediction

• GENSCAN is a widely used ab initio gene prediction tool for eukaryotic genomes
• It employs a GHMM with states for exons, introns, and intergenic regions, as well as signals like start codons and splice sites
• GENSCAN incorporates various biological features, such as codon usage, CpG islands, and promoter elements
• It can predict complete gene structures, including multiple exons and alternative splicing events

Glimmer for prokaryotic gene prediction

• Glimmer (Gene Locator and Interpolated Markov ModelER) is designed for gene prediction in prokaryotic genomes
• It uses interpolated Markov models (IMMs) to capture the variable-order dependencies in coding and non-coding regions
• Glimmer employs a two-phase approach: initial prediction of coding regions followed by a refinement step using IMMs
• It has been successfully applied to various bacterial and archaeal genomes and can handle short coding sequences

Comparison of ab initio tools

• Different ab initio tools have their strengths and weaknesses depending on the target genome and the specific biological features they model
• GENSCAN and Glimmer are optimized for eukaryotic and prokaryotic genomes, respectively, considering their distinct gene structures
• Some tools, like AUGUSTUS and GeneMark, offer flexibility in training on specific datasets or incorporating external evidence
• Comparative evaluations help assess the performance and suitability of different tools for a given genome annotation task

Evaluating gene prediction performance

Sensitivity vs specificity

• Sensitivity (recall) measures the proportion of true positive predictions out of all actual positives (TP / (TP + FN))
• Specificity measures the proportion of true negative predictions out of all actual negatives (TN / (TN + FP))
• A balance between sensitivity and specificity is desired, as increasing one may come at the cost of the other
• The F1 score, the harmonic mean of precision and recall, provides a single metric for overall performance

Exon-, transcript-, and gene-level accuracy

• Exon-level accuracy assesses the correctness of predicted exon boundaries compared to the actual exon structures
• Transcript-level accuracy evaluates the predicted splicing patterns and the agreement with the true transcript variants
• Gene-level accuracy measures the overall correctness of predicted gene structures, including the number and orientation of genes
• Different levels of accuracy provide insights into the strengths and weaknesses of gene prediction methods

Benchmarking on gold standard annotations

• Benchmarking datasets with high-quality, manually curated gene annotations serve as a gold standard for evaluation
• Datasets like ENCODE, RefSeq, and GENCODE provide trusted annotations for various model organisms
• Predicted gene structures are compared against the benchmark annotations to compute performance metrics
• Regularly updated benchmarking datasets incorporate new experimental evidence and improve the reliability of evaluations

Challenges and limitations

Pseudogenes and non-coding RNA genes

• Pseudogenes, non-functional gene copies, can be mistakenly predicted as protein-coding genes due to sequence similarity
• Non-coding RNA genes (microRNAs, lncRNAs) lack typical coding features and are often missed by ab initio gene predictors
• Distinguishing pseudogenes and non-coding RNA genes requires additional computational methods and experimental validation
• Incorporating RNA-seq data and comparative genomics can help identify and filter out pseudogenes and predict non-coding RNA genes

Alternative splicing and isoforms

• Alternative splicing generates multiple transcript isoforms from a single gene locus, increasing proteome diversity
• Ab initio gene predictors often struggle to accurately predict all possible isoforms and their relative abundances
• Isoform prediction requires the integration of RNA-seq data and machine learning approaches to model splicing patterns
• Challenges include identifying rare isoforms, predicting microexons, and resolving complex alternative splicing events

Improving predictions with homology

• Homology-based gene prediction leverages sequence conservation across related species to refine ab initio predictions
• Protein sequence alignments and synteny information can guide the identification of exon-intron boundaries and gene structures
• Integrating ab initio predictions with homology evidence can improve the accuracy and completeness of gene annotations
• Challenges include handling gene duplication events, lineage-specific gene losses, and divergent sequences with limited conservation