The foundations of bioinformatics
Bioinformatics treats living things as information. A genome is a text of four letters, a protein a text of twenty, and the field is the set of methods for reading, comparing and making sense of those texts — now billions of letters at a time.
A result that depends on which software version ran on which day is not yet a result. The field rests on public archives that anyone can query, on shared file formats, and on recorded workflows. The FAIR principles of 2016 — findable, accessible, interoperable, reusable — put the expectation into words.
Genes are predicted from raw sequence, function from similarity, and structure from sequence. In 2020, at the CASP14 assessment, AlphaFold 2 predicted most of the test structures with accuracy comparable to experiment; computational protein design and structure prediction were recognised with the 2024 Nobel Prize in Chemistry.
Search a large enough database and something will always match. Karlin and Altschul gave, in 1990, the statistics behind the E-value: how many hits this good would be expected by chance. When thousands of genes are tested at once, the false discovery rate of Benjamini and Hochberg (1995) plays the same part.
Sequencing machines read short pieces, not chromosomes. Assembly rebuilds the whole from overlaps between the pieces; mapping places each piece on a genome that is already known. Both are problems about strings and graphs, and both had to be re-solved when read counts rose from thousands to billions.
Two sequences are usually alike because they descend from a common ancestor — they are homologous. That is why a gene studied in yeast says something about its human counterpart. Margaret Dayhoff’s substitution matrices of 1978 measured which changes evolution tolerates; Joseph Felsenstein’s likelihood method of 1981 turned aligned sequences into family trees.
Lining two sequences up, allowing for substitutions and gaps, is the basic operation of the field. Needleman and Wunsch solved it exactly in 1970 with dynamic programming; Smith and Waterman gave the local version in 1981. BLAST, in 1990, traded a little sensitivity for the speed to search whole databases.
Frederick Sanger sequenced the two chains of insulin in the early 1950s and in 1955 worked out how they are joined, showing that a protein is a definite string of residues. In 1965 Margaret Dayhoff published the first collection of protein sequences — sixty-five of them — meant to be compared. Everything since is that idea at larger scale.
Read from the bottom up. Almost every method above depends on the second idea: that two sequences can be compared, and that the comparison means something.