7.1 Shotgun assembly #
7.1.1 Reads, contigs, scaffolds #
Shotgun sequencing breaks DNA at random and infers order from overlaps. A contig is sequence the assembler will write without a gap. A scaffold orders contigs across gaps, often runs of N whose lengths are estimates from linking data. Quoting N as a base is a reporting error.
Overlap–layout–consensus suits longer reads. A de Bruijn graph suits short reads, which are too numerous to compare all against all: a k-mer is a node and a (k+1)-mer is an edge from its prefix to its suffix, and the genome is a path. Both methods fail where that path is not unique. N50 measures contiguity, not correctness. Sort contigs from long to short until their sum reaches half the assembly; that contig’s length is the N50. Lengths 400, 300, 200, 50, and 50 sum to 1,000, so the N50 is 300. A false join raises the N50 and makes the assembly worse.
7.1.2 Why repeats break short-read assemblies #
A repeat longer than the reads’ distinguishing reach collapses the graph. Copies longer than k become one tangle. Raising k separates more repeats until errors and shallow coverage break true k-mers. An error and a real polymorphism both make a spur or a bubble, and pruning removes both.
Paired ends record a distance. If the repeat is shorter than the fragment, the two reads can land in unique sequence and name the flanks. If every fragment is shorter than the repeat, they cannot. Extra short reads polish unique sequence once it is well covered. They do not bridge a repeat longer than the library.
7.1.3 What long reads and proximity data each contribute #
A long read that crosses a repeat, with unique sequence on both sides, carries the bases of that path. A pair carries only a distance. An array longer than the reads still collapses. Homopolymer length remains a systematic weakness of several long-read chemistries, and averaging a systematic error does not remove it.
Hi-C-style data record which loci were close in the nucleus. Contacts fall off with genomic distance and are far more common within a chromosome than between chromosomes, which orders and orients scaffolds and can flag a bad join. They do not sequence across a repeat. Long reads contribute spanned bases; proximity data contribute order. These notes state what the data are, not how a library is built.
7.1.4 Coverage, as a model #
Coverage c = NL/G is redundancy, not the fraction of the genome seen. If placements are random and independent, the unseen fraction is about e^(−c). That is the Lander–Waterman picture in one expression.
| Coverage c | Unseen fraction, about e^(−c) | Reading |
|---|---|---|
| 1 | 0.37 | About a third of bases missed |
| 5 | 0.0067 | About 0.7% missed, if placement is random |
| 10 | about 5 × 10⁻⁵ | Almost every base hit, in the model |
| 30 | negligible in the model | Further depth serves errors and alleles |
GC bias and repeats break the assumptions, so one number mixes unique sequence with sequence that cannot be placed. A tiny unseen fraction can sit beside an assembly still broken at repeats. Variant calling wants a different depth: a heterozygote is on about half the reads, and errors must be outvoted. A few tens of times coverage is a common practical range for germline short-read calls. It is not a consequence of e^(−c).
7.2 Finding the genes #
7.2.1 Bacteria, where open reading frames are the easy part #
A bacterial gene is often a long open reading frame. If codons were equal, a stop (3 of 64) would appear about every 64/3 ≈ 21 codons. The chance that 100 given codons contain no stop is (61/64)¹⁰⁰, about 0.008. Many windows exist, so one such ORF is only a candidate; an ORF of several hundred codons usually is a gene. Starts are mostly ATG, sometimes GTG or TTG. A Shine–Dalgarno sequence supports a start. It does not define one.
Stops are AT-rich, so GC-rich DNA grows long chance ORFs and length becomes a weaker call. Short real genes look like that background. Function is then transferred by search and by domain models, with the Chapter 2 warning intact. “Hypothetical protein” means the transfer failed, not that the chain does nothing.
7.2.2 Eukaryotes, where the predictors disagree #
Introns mean that stop-to-stop on the chromosome is not the gene. Splice signals are short: GT–AG is real and far too common to place exons by itself. Ab initio finders combine those weak signals with statistics trained on known genes. GENSCAN and AUGUSTUS are classical examples of the class, not software versions to review.
They disagree because the signals are weak, because alternative splicing means more than one true transcript, and because a model trained on human genes is not an oracle for a distant species. An unsupported prediction is a hypothesis about the gene set, not a measurement of it.
7.2.3 Evidence-based annotation #
Protein alignments tend to cover exons and skip introns, and a frameshift in that alignment is an alarm. Transcript evidence, from cDNA or RNA-seq, places exons directly and can include untranslated regions. A gene silent in the sample is missing from that evidence, not from the genome.
Structural Annotation Flowchart
Where evidence and an ab initio call conflict, the evidence should win, and the release should say which was used. Alternative transcripts should stay alternative. “The annotation” is a dated set of those decisions, not a fact the chromosome contains.
7.2.4 Browsers #
The UCSC Genome Browser and Ensembl draw tracks on one assembly: sequence, gene models, RNA-seq, repeats, variants, conservation. Tracks need not agree. Two gene models are two claims. A conservation track is a computation, and a coverage peak is a count, not a protein.
The browser is a way to look, not a source of truth. A picture without the assembly accession, the track name, and the date cannot be checked. The useful click is the one that opens the alignment or the evidence code.
7.3 Key terms, summary, and self-check #
Key terms. Shotgun; contig; scaffold; N50; de Bruijn graph; k-mer; paired end; long read; Hi-C; coverage; Lander–Waterman; open reading frame; ab initio gene finding; evidence-based annotation; genome browser.
Summary. Contigs are unbroken sequence; scaffolds order them across estimated gaps; N50 measures length and rises when a join is wrong. Repeats longer than k or the fragment collapse short-read assemblies. Long reads contribute bases they span; proximity data contribute neighbours that order chromosomes. The unseen fraction under a random model is about e^(−c), and real libraries violate the model. Bacterial genes are often long ORFs, less clearly so in GC-rich DNA. Eukaryotic predictors disagree, and protein or transcript evidence should outrank an unsupported call. A browser displays claims. It does not settle them.
Self-check.
1. Joining those five contigs into one piece of 1,000 moves N50 from 300 to 1,000. Why is that not progress if the join is false?
2. Why does coverage of 200× not resolve a repeat longer than the fragment?
3. What does a long read add that a Hi-C contact does not, and the reverse?
4. About what fraction of bases does the Poisson model miss at coverage 5, and why is a eukaryotic short-read assembly at 5× worse than that figure?
5. Only one of two gene tracks has RNA-seq reads split on its splice junctions. Which claim has the direct evidence?
—
—
—