Sequencing and Genomes

Two drafts of the human genome were published in February 2001. The first assembly in which every human chromosome but one ran from end to end without a gap was published in 2022.

Why the twenty-year wait, when the machines got faster every year? Speed was never the bottleneck. No instrument reads a chromosome in one pass. Every method reads pieces, and a genome is only as good as the way the pieces are put back together. This lesson is about reading the pieces, counting them, and the one feature of genomes that defeats short pieces: repeats.

Reading DNA one base at a time: Sanger’s stopped copies

Imagine photocopying a long sentence a thousand times on a copier that, at random, quits after some letter and highlights the letter it quit on. Sort the copies by length and read the highlighted letters in order, and you have the sentence.

Sanger sequencing does this with a DNA polymerase. You supply a single-stranded template, a short primer (a piece of DNA that pairs with the template and gives the polymerase a 3’ end to extend), the four ordinary deoxynucleotides, and a small amount of dideoxynucleotides (ddNTPs). A ddNTP lacks the 3’ hydroxyl that the next nucleotide must bond to, the chemistry you met in nucleotide metabolism. The polymerase adds it opposite its partner base, and that copy can grow no further. Across millions of template molecules, every possible stopping point is represented.

Frederick Sanger, Steve Nicklen and Alan Coulson published the method in 1977, using DNA from the phage phiX174. In the classical version, four tubes each held one kind of ddNTP, and the radioactive fragments ran in four lanes of a gel that separates pieces differing by a single nucleotide. Later versions put a different fluorescent dye on each of the four terminators, so one reaction holds all four, and a capillary sorts the fragments by size while a detector reads the colour of each. Reading from the shortest fragment to the longest gives the new strand, 5’ to 3’; the template is its reverse complement.

Sanger reads are accurate, but each reaction reads one region, commonly up to about 800 to 1,000 bases, and the first few dozen bases are often poor. The sequence in the 2001 drafts came mostly from this chemistry, run on many instruments at once.

Reading millions at once, short and long

Short-read sequencing reads hundreds of millions of fragments in parallel. Genomic DNA is broken into fragments a few hundred bases long and fitted with adapters. In the method David Bentley and colleagues described in 2008, which became the Illumina platform, each fragment is copied in place on a glass flow cell into a cluster of identical copies. Then the polymerase adds one nucleotide per cycle. Each nucleotide carries a dye and a reversible terminator, a removable block on the 3’ end. The machine images every cluster, reads the colour, then cleaves off the dye and the block to regenerate the 3’ hydroxyl for the next cycle.

This is not Sanger run in parallel. Sanger stops chains for good and reads lengths; sequencing by synthesis pauses every chain at each base and reads colours, cycle by cycle. Each fragment is often read from both ends, a paired-end read, so you know two short sequences and roughly how far apart they lie. Reads are short (35 bases from each end in the 2008 human genome; commonly 100 to 150 now), and the errors in the study of one such instrument were mostly substitutions.

Long-read methods read single molecules. In PacBio instruments one polymerase sits at the bottom of a tiny well and copies a circular template, with the dye on the nucleotide’s terminal phosphate so it is released as the base joins; going round the circle many times gives HiFi reads, in a 2019 study of one human genome averaging 13.5 kilobases at 99.8% accuracy, with leftover errors mostly small insertions or deletions in runs of one base. In Oxford Nanopore instruments a strand passes through a protein pore and software decodes the changing ionic current into bases; a 2018 study of human DNA reported ultra-long libraries in which half the sequence lay in reads longer than 100 kilobases. The trade-off is plain: short reads give many accurate bases per run but cannot span long repeats; long reads span them.

Coverage and assembly: the repeat problem

A read is the sequence reported from one fragment. Coverage (depth) is how many reads, on average, lie over each base:

C = N x L / G, where N is the number of reads, L the read length and G the haploid genome length

Reads are then either aligned to a reference genome, or assembled de novo, without one, by their overlaps. A de novo assembly yields contigs, continuous consensus sequences, and scaffolds, contigs placed in order and orientation with estimated gaps between them.

Worked problem. Assume reads start at random positions, all reads are usable, and the haploid human genome is G = 3.1 x 10^9 bases (an estimate rounded for the calculation).

(a) How many 150-base reads give 30-fold mean coverage?

N = C x G / L = 30 x 3.1 x 10^9 / 150 = 9.3 x 10^10 / 150 = 6.2 x 10^8 reads

That is 620 million reads, or 310 million read pairs of 2 x 150 bases, and 93 billion bases. Check: 6.2 x 10^8 x 150 = 9.3 x 10^10, and dividing by 3.1 x 10^9 gives 30.

(b) How much of the genome gets no read at all? With random starts, the number of reads over a base follows a Poisson distribution with mean C, so the fraction with zero reads is e^-C.

C = 1: e^-1 = 0.37, so 37% unread

C = 5: e^-5 = 0.0067, so 0.0067 x 3.1 x 10^9 = about 21 million bases unread

C = 30: e^-30 is about 9 x 10^-14, far less than one base

So 30-fold is a mean, not a guarantee, and real libraries cover some regions, such as very GC-rich ones, less evenly than this model assumes. This Poisson treatment follows Eric Lander and Michael Waterman’s 1988 analysis of random clones.

(c) Now give the genome two identical copies of a 6-kilobase repeat. A 150-base read from inside the repeat fits both copies equally well. Reads at its edges are unique, but no read, and no pair from a 400-base fragment, reaches across 6 kilobases to the far edge. The assembler cannot tell which flanking sequences belong together, so it collapses the two copies into one or breaks the contig. Deeper coverage adds more reads of the same ambiguous kind. A 15-kilobase read, by contrast, can hold a whole 6-kilobase copy with several kilobases of unique sequence on each side, and that pins the copy in place. Some human-specific gene duplications described in the human brain expansion lesson sit in long near-identical repeats like these, one reason they were hard to study.

The human reference: gaps, a finished genome and a pangenome

The 2001 papers, from the public International Human Genome Sequencing Consortium and from Celera Genomics, were drafts. The consortium’s “finished” sequence of 2004 covered about 99% of the euchromatin (the gene-rich, less tightly packed part of the genome) and still had 341 gaps, many associated with segmental duplications; centromeres and other heterochromatin were not part of its target. The same paper estimated 20,000 to 25,000 protein-coding genes, far fewer than the 2001 estimate. The reference most laboratories use today, GRCh38, descends from it, and it is a mosaic of DNA clones from several people, not the genome of any one person.

The gaps were the repeats from the worked problem, at chromosome scale. The Telomere-to-Telomere Consortium closed them using the CHM13 cell line, whose two copies of each chromosome are nearly identical, and a combination of PacBio HiFi and Oxford Nanopore ultra-long reads. Their 2022 assembly, T2T-CHM13, is 3.055 billion bases and gapless for every chromosome except Y, which CHM13 lacks. A complete Y from a different person, HG002, followed in 2023.

One complete genome still represents one person. In 2023 the Human Pangenome Reference Consortium released a draft pangenome: 47 people’s genomes, each with its two chromosome copies assembled separately, joined into a graph in which each person’s copy is a path. It added 119 million bases of variable euchromatic sequence missing from GRCh38.

What varies between two genomes, and why the reference matters

Differences come in sizes. A single-nucleotide variant (SNV) changes one base. A small indel inserts or deletes a few bases. A structural variant is larger, by a common convention 50 bases or more, and includes deletions, duplications, inversions, insertions and translocations; when a deletion or duplication changes the number of copies of a segment, it is a copy-number variant. The chromosome-scale versions are in chromosome changes, and how such changes arise is in mutation.

In the 1000 Genomes Project’s 2015 analysis of 2,504 people from 26 populations, a typical genome differed from the reference at 4.1 to 5.0 million sites, and structural variants affected more bases than SNVs and small indels together. Every one of those calls is a difference from a reference. Where the reference carries a rare allele, most people look “variant” there; where a sequence is missing from the reference, reads from it have nowhere to align. In the Human Pangenome Reference Consortium’s benchmarks, short-read workflows using the pangenome made 34% fewer small-variant errors and found 104% more structural variants per haplotype than workflows using GRCh38.

Short reads detect large variants indirectly, from depth (copy number changes the number of reads), discordant pairs (two ends of a fragment mapping farther apart than the fragment could be) and split reads (one read aligning in two pieces across a junction).

Design problem

Short paired-end reads from a patient sample are aligned to the reference. The fragments are normally 400 ± 50 bases long. Depth is 30-fold across a chromosome segment but 15-fold across one 10-kilobase stretch within it. Forty read pairs map with their two ends 10.4 kilobases apart across that stretch, and several reads split, aligning partly to each edge. What does this pattern show, and how would you confirm it?