Describe genome annotation — Biology notes

Core Concept

Genome annotation is the systematic process of identifying and labeling all functional elements in a genome sequence—including genes, regulatory elements, repeat sequences, and other biologically meaningful features.

WHY do we need this? A raw genome sequence has zero context—it's like having all the letters of an encyclopedia dumped in random order. Annotation transforms sequence data into biological knowledge.

HOW does it work? Through computational prediction algorithms combined with experimental validation.

The Annotation Pipeline

1. Structural Annotation (Finding the genes)

Formula Gene Prediction Approaches

Ab initio (de novo) prediction:

Uses statistical models of gene structure (Hidden Markov Models, neural networks)
Looks for signals: start codons (ATG), stop codons (TAA, TAG, TGA), splice sites (GT-AG), promoter motifs
Derivation of probability: For a sequence $S$ , we want $P(\text{gene}|S)$ $P(\text{gene}|S) = \frac{P(S|\text{gene}) \cdot P(\text{gene})}{P(S)}$

Where:

$P(S|\text{gene})$ = likelihood the sequence looks like known genes (learned from training data)
$P(\text{gene})$ = prior probability (roughly gene density in the genome)
$P(S)$ = normalizing constant

WHY this works: Genes have non-random patterns—codon usage bias, length distributions, conserved splice sites. The model learns these patterns from verified genes.

Homology-based prediction:

Compare against known genes from other organisms using BLAST or similar alignment
If you find 80% sequence identity with a human insulin gene, it's probably an insulin gene $\text{Similarity Score} = \frac{\text{Number of matches}}{\text{Alignment length}} \times 100\%$

Evidence-based prediction:

Use actual experimental data: RNA-seq (shows what's transcribed), proteomics (shows what's translated)
This is ground truth—if you sequence mRNA and it aligns to a genomic region, that region is definitely transcribed

2. Functional Annotation (What do the genes DO?)

Worked example Example 1: Annotating a Bacterial Gene

Scenario: You've identified an open reading frame (ORF) of 1,200 bp in E. coli.

Step 1 - Structural:

Found start codon ATG at position 1000 (first base of the ORF)
Found stop codon TAA spanning positions 2197–2199 (last three bases of the ORF)
ORF length = 2199 − 1000 + 1 = 1200 bp → 400 codons → 399 amino acids (stop codon codes no residue)

WHY this step? We're confirming this has the basic structure of a gene: a start codon, a contiguous in-frame reading region, and a proper 3-base stop codon (TAA/TAG/TGA). The coordinates must be consistent — start position + length − 1 = stop end position.

Step 2 - Homology Search:

BLAST against NCBI database
Top hit: E. coli lactose permease (lacY), 95% identity WHY this step? Sequence similarity implies functional similarity (homologs usually share function).

Step 3 - Domain Search:

Pfam search finds "Major Facilitator Superfamily" domain
This domain is found in membrane transport proteins

WHY this step? Protein domains are functional units—finding known domains tells us molecular function.

Step 4 - Pathway Analysis:

KEGG database shows this gene is in the "Galactose metabolism" pathway
Interacts with lacZ (beta-galactosidase) and lacA (transacetylase) WHY this step? Function emerges from context—knowing pathway placement gives biological meaning.

Final Annotation:

Gene: lacY_homolog
Product: Lactose permease
Function: Transmembrane transport of lactose
Pathway: Galactose metabolism (KEGG map00052)
GO Terms: GO:0015153 (galactose transmembrane transporter activity)

Worked example Example 2: Annotating a Human Gene with Alternative Splicing

Scenario: You've found a 50,000 bp genomic region with RNA-seq evidence.

Step 1 - Exon Detection:

RNA-seq shows reads mapping to 5 distinct regions (exons)
Splice junction reads confirm: Exon1-Exon2, Exon1-Exon3, Exon2-Exon4, Exon3-Exon4, Exon4-Exon5

WHY this matters? This gene produces multiple isoforms through alternative splicing. Step 2 - Isoform Reconstruction:

Isoform A: E1-E2-E4-E5 (skips E3)
Isoform B: E1-E3-E4-E5 (skips E2) Step 3 - Functional Prediction:
Isoform A: 350 amino acids, contains kinase domain → probably a signaling enzyme
Isoform B: 280 amino acids, kinase domain truncated → probably a dominant-negative regulator

WHY different functions? Alternative splicing often removes or adds functional domains, creating proteins with different or opposing activities.

Annotation Challenge: Must record BOTH isoforms with their tissue-specific expression patterns.

Annotation Quality Metrics

Formula Sensitivity, Specificity and Precision

Gene prediction tools are evaluated with a confusion matrix of four counts: TP, FP, FN, TN. Three key metrics:

$\text{Sensitivity (Recall)} = \frac{TP}{TP + FN}$

$\text{Specificity} = \frac{TN}{TN + FP}$

$\text{Precision (Positive Predictive Value)} = \frac{TP}{TP + FP}$

Where:

TP = True Positives (correctly identified genes)
FP = False Positives (predicted genes that aren't real)
FN = False Negatives (real genes that were missed)
TN = True Negatives (non-gene regions correctly called non-gene)

WHY three metrics?

Sensitivity asks: of all the real genes, how many did we catch?
Specificity asks: of all the non-gene regions, how many did we correctly leave out?
Precision asks: of all the things we called genes, how many are actually genes?

Careful — a common confusion: TP/(TP+FP) is precision, NOT specificity. Specificity uses true negatives: TN/(TN+FP). In gene finding, precision is often reported because true negatives (all non-coding positions) are enormous and hard to define.

WHY the sensitivity–precision trade-off?

High sensitivity (catching all genes) → more false positives → lower precision
High precision (only confident calls) → miss rare or unusual genes → lower sensitivity

Typical tools achieve ~90% sensitivity and ~85% precision on well-studied organisms.

Types of Annotation Challenges

Challenge 1: Non-coding RNA genes

WHAT: Genes that produce functional RNA instead of protein (tRNA, rRNA, miRNA, lncRNA) WHY hard: No open reading frame, no start/stop codons—must rely on secondary structure or conservation

Challenge 2: Pseudogenes

WHAT: Gene-like sequences that have lost function (often have stop codons or frameshifts) WHY important: Can be mistaken for real genes; understanding pseudogenes reveals evolutionary history

Challenge 3: Overlapping genes

WHAT: One gene encoded within another (opposite strand or different reading frame) WHY hard: Standard gene-finders assume non-overlapping genes; requires special detection

Common mistake Common Annotation Errors

Mistake 1: "Every ORF is a gene" Why it feels right: Open reading frames (start → stop with no interruptions) look like genes.

Why it's wrong: In random DNA, stop codons appear frequently — the mean run between stop codons is only about 63 bp (≈ 21 codons), because 3 of the 64 codons are stops. So short ORFs occur constantly, but genuinely long ORFs (>300 bp) are statistically rare in random sequence. A short ORF is essentially expected by chance and means almost nothing on its own.

The fix: Require additional evidence—homology to known genes, RNA-seq support, conservation across species, and a length longer than random expectation. An ORF is a candidate, not a confirmation.

Mistake 2: "High BLAST similarity = identical function"

Why it feels right: If two proteins are 80% identical, they probably do the same thing. Why it's wrong: Small sequence differences can change function dramatically, and even fairly low identity can hide within a shared fold. Example: rhodopsin (light sensor) and olfactory receptors both belong to the class-A GPCR family and share only about 20–25% identity, yet have completely different ligands and functions. Similar fold ≠ identical function.

The fix: Use similarity as starting hypothesis, then check:

Are key catalytic residues conserved?
Does domain architecture match?
Is expression pattern similar?

Mistake 3: "No known homolog = not a gene" Why it feels right: If BLAST finds nothing, maybe it's junk DNA.

Why it's wrong: Every gene family had a "first member" with no homologs. Rapidly evolving genes (like immune system or toxin genes) often have no detectable homologs. The fix: Use multiple lines of evidence—RNA-seq, proteomics, evolutionary conservation of synteny (gene order), purifying selection signatures.

The Human Genome Annotation Saga

Automated vs Manual Annotation

Aspect	Automated	Manual (Curation)
Speed	Fast (millions of genes/day)	Slow (hours per gene)
Consistency	High	Variable
Accuracy	~85-90%	~95-99%
Cost	Low	High (expert time)
Best for	First-pass annotation	Model organisms, clinical genes

Reality: Use automated for genome-wide coverage, manual curation for important genes (disease-related, drug targets).

Recall Explain to a 12-year-old

Imagine you just received a massive LEGO instruction manual, but it's all in code—just 1s and 0s. Genome annotation is like having a computer (and smart people) go through and label each part: "These numbers mean build a car," "These mean build a house," "These are the instructions for the color scheme."

But here's the tricky part: sometimes the instructions overlap, or they're written backwards, or some instructions are broken copies of other instructions. The annotation team has to figure out what's a real instruction, what's a mistake, and what each instruction actually builds.

When scientists finish annotating a genome, they know: "This organism has about 20,000 different protein-building instruction sets (genes), and we're confident about what most of them build." Without annotation, the genome is useless data. With annotation, it's a blueprint for life.

Connections

6.1.01-DNA-sequencing-methods - Annotation requires high-quality sequence first
6.1.05-Comparative-genomics - Cross-species comparison improves annotation accuracy
Gene-expression-profiling - RNA-seq provides direct evidence for transcribed regions
Protein-structure-prediction - Functional annotation benefits from knowing 3D structure
BLAST-and-sequence-alignment - Core tool for homology-based annotation
Hidden-Markov-Models - Statistical framework for ab initio gene prediction

#flashcards/biology

What is genome annotation? :: The systematic process of identifying and labeling functional elements in a genome sequence, including genes, regulatory elements, and repeat sequences.

What are the two main types of genome annotation?

Structural annotation (identifying locations/structures of genes, exons, introns) and functional annotation (assigning biological functions to identified genes).

What is ab initio gene prediction?

A computational method that uses statistical models to predict genes based on patterns in the sequence itself, without relying on comparison to known genes.

What is homology-based annotation?

Identifying genes by comparing the genome sequence to databases of known genes from other organisms using tools like BLAST.

Why can't we trust every open reading frame (ORF) as a real gene?

In random DNA the mean run between stop codons is only ~63 bp, so short ORFs arise constantly by chance; only long ORFs (>300 bp) with supporting evidence (RNA-seq, homology, conservation) are likely real.

What is the difference between specificity and precision in gene prediction?

Specificity = TN/(TN+FP), the fraction of non-gene regions correctly excluded. Precision (PPV) = TP/(TP+FP), the fraction of predicted genes that are real. They are NOT the same.

What is sensitivity (recall) in gene prediction?

Sensitivity = TP/(TP+FN), the fraction of real genes that were correctly identified.

What makes non-coding RNA genes hard to annotate?

They lack open reading frames and translation signals, so annotation must rely on secondary structure predictions or evolutionary conservation patterns.

What is a pseudogene?

A gene-like sequence that has lost function through mutations (stop codons, frameshifts), often resembling a real gene but not producing functional product.

Why has the human gene count decreased over time since 2001?

Better algorithms and experimental validation revealed that many initial predictions were pseudogenes, annotation errors, or non-coding elements rather than true protein-coding genes.

What types of experimental evidence validate gene annotations?

RNA-seq (shows transcription), proteomics (shows translation), ChIP-seq (shows regulatory binding), and evolutionary conservation across species.

Concept Map

Hinglish (regional understanding)

Intuition Hinglish mein samjho

Dekho, genome annotation ka core idea bahut simple hai — socho tumhare paas ek gigantic instruction manual hai jo kisi alien language mein likha hai, na koi spaces, na punctuation, na chapter names. Bas millions of A, T, G, C letters ki ek continuous line. Ab is raw sequence ka koi matlab tab tak nahi jab tak tum usme mark na karo ki "yeh part protein banata hai," "yeh ek control switch hai," "yeh part abhi tak samajh nahi aaya." Yehi kaam annotation karta hai — sequence ko biological knowledge mein convert karna. Isliye yeh itna important hai, kyunki bina annotation ke genome sirf random letters ka pile hai, uska koi scientific use nahi.

Ab yeh process do main hisso mein hota hai. Pehla hai structural annotation — matlab genes kahan hain, exons-introns kahan hain, promoters kahan hain, yeh sab locate karna. Iske liye scientists teen tareeke use karte hain: ab initio prediction jisme statistical models (jaise Hidden Markov Models) start codon (ATG), stop codons (TAA/TAG/TGA), aur splice sites (GT-AG) jaise signals dhoondte hain; homology-based prediction jisme dusre organisms ke known genes se BLAST kar ke compare karte hain (agar 80% match human insulin gene se hai, toh probably insulin hi hoga); aur evidence-based prediction jisme actual RNA-seq ya proteomics data use karke confirm karte hain ki region sach mein transcribe ho raha hai — yeh sabse solid ground truth hota hai.

Dusra hissa hai functional annotation — matlab gene karta kya hai? Sirf location pata karna kaafi nahi, humein yeh bhi jaanna hai ki yeh gene kaunsa protein banata hai aur kis pathway mein kaam karta hai. Jaise example mein dekha — ek ORF milta hai, phir BLAST se pata chalta hai yeh lacY (lactose permease) se 95% match karta hai, phir Pfam se domain milta hai jo membrane transport se related hai, aur KEGG se pura pathway samajh aata hai. Yaad rakho, yeh probability wala formula $P(\text{gene}|S)$ isliye kaam karta hai kyunki genes ke patterns random nahi hote — codon usage bias, length distribution, conserved splice sites — model in patterns ko verified genes se seekh leta hai. Toh basically annotation matlab meaningless letters ko meaningful biology mein translate karna, aur yeh modern genomics ki foundation hai.

Test yourself — Genomics

Connections