Explain gene prediction methods

WHY do we need gene prediction at all?

WHAT: Sequencing gives you raw DNA — billions of letters. But raw sequence does not tell you which stretches actually code for proteins or functional RNA.

WHY it's hard:

• There are no obvious markers separating a gene from junk DNA.
• In eukaryotes, genes are interrupted: coding pieces (exons) are broken up by non-coding pieces (introns) that get spliced out.
• Any DNA stretch has 6 reading frames (3 forward, 3 reverse), and most are meaningless.

HOW we attack it: We look for statistical signals and sequence signals that real genes leave behind, then combine them.

The two big families of methods

1. Ab initio (intrinsic) methods

Use only the query sequence and statistical models of what genes look like.

2. Homology / evidence-based (extrinsic) methods

Compare the sequence to known genes, proteins, or expressed sequences (mRNA/EST/RNA-seq).

Sequence signals we hunt for (content vs. signal)

Building block 1 — Open Reading Frames (ORFs)

WHY ORFs matter: In prokaryotes, genes are (mostly) continuous ORFs, so ORF-finding alone works well. In eukaryotes, introns break the ORF, so ORFs are only part of the story.

Deriving: how many codons, how long an ORF?

DNA is read in triplets. If an ORF spans $n$ nucleotides (including start & stop), the number of codons is:

$C = \frac{n}{3}$

The number of amino acids in the protein (start codon → Met is kept, stop codon → no amino acid):

$L_{aa} = \frac{n}{3} - 1$

Deriving: probability a random codon is a stop

There are $4^3 = 64$ codons. Three of them (TAA, TAG, TGA) are stops.

$P(\text{stop}) = \frac{3}{64} \approx 0.047$

Expected number of codons before hitting a stop (geometric distribution mean):

$E[\text{codons before stop}] = \frac{1}{P(\text{stop})} = \frac{64}{3} \approx 21$

Building block 2 — Markov models & GHMMs (the engine of modern predictors)

Deriving a k-th order Markov probability

A sequence $S = s_1 s_2 \dots s_n$. Under a $k$-th order Markov model:

$P(S) = P(s_1 \dots s_k)\prod_{i=k+1}^{n} P(s_i \mid s_{i-1}, \dots, s_{i-k})$

Why this product form? We apply the chain rule of probability and then make the Markov assumption: each base depends only on the previous $k$ bases, not the entire history. This makes the huge joint probability tractable (only need $4^{k+1}$ parameters).

We then compute a log-likelihood ratio score comparing a coding model vs a non-coding (null) model:

$S_{LLR} = \log \frac{P(\text{sequence} \mid \text{coding})}{P(\text{sequence} \mid \text{non-coding})}$

GHMM (Generalized Hidden Markov Model): tools like GENSCAN, AUGUSTUS, GeneMark, Glimmer use HMMs where hidden states = {exon, intron, intergenic, 5'UTR, splice site, ...}. The predictor finds the most probable state path (via the Viterbi algorithm) that explains the observed DNA — that path is the predicted gene structure.

Homology-based methods (borrowing the dictionary)

• BLAST/BLAT the sequence against protein/nucleotide databases → conserved regions likely code.
• EST / cDNA / RNA-seq alignment: expressed sequences map directly onto exons, revealing exon boundaries and splicing experimentally.
• Tools: GeneWise, Exonerate, spliced aligners (STAR, HISAT).

Prokaryotes vs. Eukaryotes

Worked examples

Common mistakes (Steel-man + fix)

Flashcards

Connections

• Sequence Alignment (BLAST) — homology-based prediction relies on it
• Hidden Markov Models — the statistical engine of GENSCAN/AUGUSTUS
• Central Dogma — transcription/splicing/translation define what a "gene" is
• Genetic Code & Codons — source of codon bias and stop-codon statistics
• RNA-seq & Transcriptomics — expression evidence for exon boundaries
• Genome Annotation Pipelines — MAKER/BRAKER combine all methods
• Prokaryote vs Eukaryote Genome Organization — why methods differ

Concept Map

Hinglish (regional understanding)