In April 2003 a headline said the human genome was finished. In March 2022 another said it was finished, again. Both were true. The first was a reference with gaps that held about 92% of the sequence; the second, 19 years later, filled in the rest. How can a text be "complete" twice? And how do you read three billion letters when your machine can only read a few hundred at a time?
The answers are the core of bioinformatics: biology treated as information. A graduate sees a genome as a very long text with typos, repeats, copies and relatives, and sees biology questions as questions about strings: where does this read go, how similar are these two, which letters travel with a disease, and what shape does this sequence fold into? This piece walks through the five ideas that every course in the subject builds on.
DNA is a four-letter text (A, C, G, T). A human haploid genome is about 3.1 billion letters, all copied in nearly every cell. No machine reads it end to end. Instead it is broken into pieces, every piece is read, and a computer stitches them back together by finding overlaps. This is shotgun sequencing, the method behind the genomes since the 1990s: in 1977 Frederick Sanger's group read the 5,386 letters of a small virus, phiX174; in 1995 the 1.8 million letters of the bacterium Haemophilus influenzae were the first free-living organism's genome; the Human Genome Project ran from 1990 to 2003, cost around $3 billion, and delivered a rough draft in 2000 and 2001 and a declared finish in April 2003. Since then the cost of sequencing has fallen far faster than the cost of computing did under Moore's law, which is why genomes of thousands of species and millions of people exist.
Two numbers decide whether the assembly works. Coverage c is the average number of times each letter is read: reads × read length ÷ genome length. Reads land at random, so by the Lander–Waterman argument the fraction of the genome that nothing ever hit is e−c.
The second number is the read length against the length of repeats. A read can only place itself if it contains something unique. The human genome is full of repeated stretches, some thousands of letters long and some, like the centromeres, millions long. A 150-letter read inside a 6,000-letter repeat has no way to know where it is. Long reads that reach out of the repeat can.
Genome 3,100 million letters. Reads = c × genome ÷ read length. Uncovered = e−c (idealised random reads, no errors).
At 30× with 150-letter reads you need about 620 million reads, around 93 billion letters, and almost nothing is missed. But a 6,000-letter repeat is still unresolved. That is exactly why the 2003 "complete" genome kept holes, and why the 2022 genome needed reads tens of thousands of letters long.
Once you have a sequence, the first thing you ask is: what does it resemble? Two similar sequences are rarely identical: letters are swapped, and some are inserted or deleted. An alignment lines two sequences up, with gaps, so that matching letters sit in the same column. Which alignment is best? Give points for matches, take points for mismatches, and take more for opening a gap, then find the arrangement with the highest total. In 1970 Saul Needleman and Christian Wunsch showed how to do that exactly, by filling in a table:
This is dynamic programming: don't search every possible alignment (there are astronomically many), solve small prefixes once and reuse them. Smith and Waterman (1981) adapted it to find the best local match, and BLAST (1990) made a fast approximate version that searches databases of billions of letters in seconds. Try your own sequences below.
Notice what the scores mean. Make gaps cheap and the program inserts gaps freely to force matches; make them dear and it prefers mismatches. There is no "true" alignment from the scores alone: the numbers encode a belief about how often evolution inserts, deletes and substitutes. Real tools use substitution matrices (BLOSUM, PAM) built from observed protein changes.
Align many species and count the differences between each pair: you get a table of distances. Humans and chimpanzees differ at roughly 1 to 1.5% of aligned letters, humans and orangutans at about 3%, humans and macaques at about 6 to 7%. A tree is a way to draw such a table. The oldest recipe is UPGMA (Sokal and Michener, 1958): repeatedly join the two closest groups and average the distances. It is simple, and it quietly assumes a molecular clock: every lineage changes at the same rate. When rates differ, it goes wrong. In 1987 Naruya Saitou and Masatoshi Nei introduced neighbor-joining, which corrects each distance by how far each taxon is from all the others, so a fast-evolving "long branch" does not fool it.
Four species A, B, C, D. The true tree pairs A with B and C with D. A and C evolve fast (long branches of length L); B and D are slow (length 1); the middle branch has length 1. (Illustrative lengths.)
With L=5 the table has AB=6, CD=6, AC=11, AD=7, BC=7, BD=3. The closest pair is B and D, at 3. UPGMA joins them, which is wrong: they are not each other's closest relatives at all, they only evolved slowly. Neighbor-joining recovers the true grouping. Below L=2 both agree. Modern methods (maximum likelihood, Bayesian inference) model rates explicitly, but the lesson stays: a method's assumptions are part of its answer.
If genomes differ by about one letter in a thousand between two people, which differences go with a trait? A genome-wide association study (GWAS) takes thousands to millions of people, records millions of variants (SNPs) and tests each one against the trait. The first big success was age-related macular degeneration in 2005. The statistics have a trap built in: test a million variants at p<0.05 and about 50,000 will "pass" by luck. The accepted threshold is therefore p<5×10−8 (roughly 0.05 divided by a million tests). A real association appears as a tall tower on a "Manhattan plot".
Simulation: 1,000 variants, 5 truly causal (marked ring), the rest null. Test statistic z ~ Normal(√(n·v/(1−v)), 1) for causal and Normal(0,1) otherwise. Fixed seed.
Slide the sample size up and causal towers rise out of the noise; with the naive threshold the null variants light up too. Three cautions every course repeats. A hit is an association, often with a neighbouring variant that is inherited with it (linkage disequilibrium), not a proven cause. Most traits are polygenic: thousands of variants each with a tiny effect, so single hits explain little. And the studies have historically been dominated by people of European ancestry, so scores built from them predict less well elsewhere: a 2019 analysis of UK Biobank-based scores found them roughly two-fold less accurate in East Asian and several-fold (about 4.5 times in one summary) less accurate in African-ancestry samples. Biobanks in Japan, China, Mexico and Africa (BioBank Japan, China Kadoorie, the Mexico City Prospective Study, H3Africa) exist partly to fix this. What "heritability" does and does not mean is in Personality and Individual Differences (35/26). Also remember the human-variation fact: most genetic variation lies within populations, not between them (Lewontin, 1972, confirmed by later genomic work). This describes research and isn't medical advice.
A protein's function depends on its folded shape, and the shape is, in principle, written in its sequence. For decades predicting it was a grand challenge. One clue is old: if two positions in a protein touch in 3-D, a mutation at one is often compensated by a change at the other, so their letters change together across species. Align hundreds of relatives and look for columns that vary in step: that is co-evolution, and you can score it with mutual information.
Toy alignment: 10 columns, 4-letter alphabet. Columns 3 and 8 co-evolve (each value of one fixes the other); the rest vary independently. Fixed seed, first n rows used.
With four or five sequences the pair you want is not even first: small samples make unrelated columns look linked. With about eight it leads, but rivals are still strong; with dozens or hundreds they fade and the true pair stands alone. That is why deep alignments mattered, and why DeepMind's AlphaFold 2 won the CASP14 competition in November 2020, scoring above 90 on the global-distance test for roughly two-thirds of targets, a level that experimentalists described as a stunning advance. It combines these alignment signals with a neural network trained on known structures (Jumper and colleagues, Nature, 2021). A public database of predictions launched in July 2021 with EMBL-EBI and grew to about 200 million structures in 2022. Demis Hassabis and John Jumper shared half of the 2024 Nobel Prize in Chemistry for it; David Baker took the other half for computational protein design.
Limits matter. A predicted structure is a hypothesis with its own confidence score, not an experiment; the model does little for proteins that move, for disordered regions, or for how a mutation changes stability. Structural biologists still do the experiments, now with a far better starting guess.
Genetic data are unlike a password. You cannot change them, and they are shared with relatives who never agreed. In 2013 Gymrek and colleagues showed that surnames could be inferred from anonymous research genomes by matching Y-chromosome patterns against public genealogy databases. In 2018 investigators identified the Golden State Killer suspect by finding distant cousins in a public genealogy database; later that year Erlich and colleagues estimated that about 60% of searches for people of European descent in the US would already find a third cousin or closer in the consumer databases of the day, and that coverage would soon be near-universal. Genetic data are protected differently around the world: the EU's GDPR classes them as a special category needing stronger grounds; the US Genetic Information Nondiscrimination Act of 2008 covers health insurance and employment but not, for example, life insurance. When a testing company went into bankruptcy in 2025, many customers asked what happens to their data in a sale, which is a good question to ask of any provider. Indigenous communities have pressed for data sovereignty after cases such as the Havasupai blood samples, used for research they had not agreed to and settled in 2010.
A genome is not just yours. Whoever holds your cousin's data holds a fuzzy copy of yours.
Biology is now also an information science: genomes are texts read in scraps, aligned by scoring, compared as distances, searched for associations and folded by learned patterns. Each step has a model with assumptions you should be able to name, and each step involves data about real people who are related to other real people.
Try this: this week, take any word and a misspelling of it (for example "separate" and "seperate") and align them by hand: mark matches, a mismatch, and find where a gap would help. You have just run Needleman–Wunsch.
Sources. Human Genome Project: declared complete 14 April 2003, about 92% of the sequence, cost roughly $3 billion (Wikipedia summary of NHGRI and DOE records; NHGRI's sequencing-cost series, last updated 2023, graphs costs against a Moore's-law line). Telomere-to-telomere genome: Nurk et al., The complete sequence of a human genome, Science 376, 44–53 (31 March 2022; DOI 10.1126/science.abj6987), 3.055 billion letters and nearly 200 million new; the Y chromosome followed in 2023. Needleman & Wunsch 1970; Smith & Waterman 1981; Altschul et al. 1990 (BLAST); Sokal & Michener 1958 (UPGMA); Saitou & Nei 1987 (neighbor-joining); Lander & Waterman 1988 coverage argument. GWAS threshold 5×10⁻⁸ is standard; Martin et al., Nature Genetics 2019 (polygenic score accuracy by ancestry; secondary coverage gives slightly different African-ancestry ratios, so the figure here is approximate). AlphaFold: CASP14 (November 2020), Jumper et al., Nature 2021, database expansion to about 200 million structures (July 2022), 2024 Nobel Prize in Chemistry. Privacy: Gymrek et al., Science 2013; Erlich et al., Science 362, 690–694 (2018); GINA 2008; GDPR. The sequence tools and charts are teaching models computed in the page; the species distances are rounded and illustrative. This describes science and isn't medical or legal advice.