Content
Summary
Eukaryotic gene prediction is a complex task due to split gene structures and low gene density, requiring programs to identify exons, introns, and splice sites. These tools utilize ab initio, homology-based, and consensus-based algorithms, employing statistical models like Hidden Markov Models (HMMs), neural networks, and discriminant analysis. Because statistical parameters are derived from specific training datasets, most programs are organism-specific.
Intuition
Imagine trying to find a needle in a haystack, but the needle is broken into pieces and scattered; this is the challenge of eukaryotic gene prediction. Computational programs act as the gatherers, using conserved sequence features like the GT-AG rule for splice junctions and the Kozak sequence for start codons to assemble these pieces. Since the statistical patterns of DNA vary between species, these programs must be trained on data specific to the organism in question, such as human or mouse, to accurately distinguish coding regions from non-coding regions and assign reliability scores to predicted exons.
Key equations
Consensus poly-A signal motif
Consensus motif for the 5′ splice junction
Consensus motif for the 3′ splice junction
Kozak sequence flanking the start codon
Probability threshold for reliable exon prediction in GENSCAN
Derivation
- Eukaryotic genomes present challenges such as split gene structures, alternative splicing, and low gene density.
- Programs categorize algorithms into ab initio, homology-based, and consensus-based types.
- Statistical models like HMMs, neural networks, and discriminant analysis identify conserved sequence features.
- Features include hexamer frequencies, splice sites (GT-AG rule), poly-A signals, and CpG islands.
- Models are trained on organism-specific data to adjust weight parameters or separation boundaries.
- Outputs include exon candidates with probability scores, where only high-confidence predictions (e.g., P > 0.5) are considered reliable.
Worked example
A researcher uses GENSCAN to predict genes in a new vertebrate sequence. The program outputs three potential exons with probability scores of 0.3, 0.6, and 0.8. Which predictions are considered reliable, and what specific sequence feature might the program have used to identify the start of the coding region?
- GENSCAN assigns a probability score (P) to putative exons.
- The program considers only predictions with P > 0.5 as reliable.
- Comparing the scores: 0.3 is not > 0.5, 0.6 is > 0.5, and 0.8 is > 0.5.
- Therefore, the exons with scores 0.6 and 0.8 are reliable.
- To identify the start of the coding region, the program likely looked for the conserved Kozak sequence (CCGCCATGG) flanking the ATG start codon.
Answer The exons with probability scores 0.6 and 0.8 are reliable. The program likely used the Kozak sequence (CCGCCATGG) to identify the translation start codon.
Common mistakes
- Assuming QDA significantly outperforms LDA in actual gene prediction, as the expected increase in performance has not been obvious.
- Ignoring that programs are often organism-specific and require training data from the specific organism being analyzed.
- Considering all predicted exons equally reliable; only predictions with a probability score P > 0.5 in programs like GENSCAN are deemed reliable.