Bioinformatics

Bioinformatics is the part of genomics that turns biological material into data a computer can handle, and data back into biological claims.

12 frameworks, from 1965 to today · Genomics · Natural Sciences

still a working view ended · each leads to its entry below

Overview

Bioinformatics is the part of genomics that turns biological material into data a computer can handle, and data back into biological claims. Its practitioners build databases, write algorithms, and design pipelines: chains of software steps that take raw output from a sequencing machine and end in an assembled genome, a list of genetic variants, or a comparison between species. The stakes are practical. A human genome is roughly three billion chemical letters, and a single sequencing run can produce billions of short fragments of text. Without methods for storing, searching, aligning, and interpreting those strings, the genomics revolution would be an archive no one could read. The history of bioinformatics is therefore the history of how the field learned to make that archive legible—and of the arguments over which computational shortcuts are scientifically honest.

Sequences Before Sequencers: The 1960s and 1970s

The first bioinformaticians worked with protein sequences, not DNA. Frederick Sanger and colleagues published the first complete amino-acid sequence of a protein, insulin, in the 1950s, and by the 1960s a small number of sequences existed. Margaret Dayhoff, a physical chemist at the National Biomedical Research Foundation in Maryland, collected them into the Atlas of Protein Sequence and Structure, first published in 1965. The Atlas was a printed book, but it was also a database in embryo: it established the idea that sequences should be gathered, curated, and compared across organisms1. Dayhoff's group also introduced the first accepted scoring scheme for comparing sequences—a table of how often one amino acid substitutes for another in evolution—and the "accepted point mutation" unit, later renamed the percent accepted mutation (PAM) unit, for measuring evolutionary distance.

The theoretical groundwork came from population genetics. Motoo Kimura's neutral theory of molecular evolution (1968) argued that most molecular variation is not selected but drifts, which gave sequence comparison a quantitative baseline. Emile Zuckerkandl and Linus Pauling had already proposed in the early 1960s that sequences could serve as a "molecular clock." These ideas made alignment meaningful: if sequences diverge at a roughly measurable rate, then comparing them recovers history.

Alignment and the First Databases: The 1970s and 1980s

The central computational problem of the 1970s was alignment: given two sequences, find the best way to line them up, allowing insertions and deletions. Saul Needleman and Christian Wunsch published a dynamic-programming algorithm for global alignment in 1970; Temple Smith and Michael Waterman gave the local-alignment version in 1981, which finds the best-matching region within two sequences rather than forcing an end-to-end match23. Dynamic programming is a method that solves a big problem by filling a table of smaller subproblems; in alignment, the table records the best score for every prefix pair2. The algorithms were exact but slow, and the field spent the following decades making them faster and more statistically principled.

The 1980s brought the first nucleotide databases4. GenBank was established in 1982 at the Los Alamos National Laboratory, and the European Molecular Biology Laboratory (EMBL) data library followed in 1980; the DNA Data Bank of Japan (DDBJ) came later in the decade. These were not merely archives: they agreed to exchange data daily, creating the international nucleotide sequence database collaboration that still operates4. The Protein Data Bank, founded in 1971 at Brookhaven, did the same for three-dimensional structures. Public data infrastructure became a precondition for everything else—a shared reference against which any new sequence could be compared.

Searching those growing databases required heuristics. In 1985, David Lipman and William Pearson published FASTA, which finds short exact matches and extends them; in 1990, Stephen Altschul, Warren Gish, Webb Miller, Eugene Myers, and David Lipman published BLAST (Basic Local Alignment Search Tool), which trades some sensitivity for enormous speed by indexing short words. BLAST became the most-used tool in molecular biology, and its statistical framework—the E-value, the expected number of chance matches of a given score—gave biologists a way to judge whether a hit meant anything.

Reading Genomes: Assembly and the Reference Human Genome

DNA sequencing in the 1980s and 1990s produced reads of a few hundred to a thousand letters. Assembling a whole genome from such fragments is a puzzle: overlap the reads, resolve repeats, and order the result1. Two strategies competed. The hierarchical (clone-by-clone) approach mapped large DNA fragments onto chromosomes first, then sequenced each fragment—the strategy of the publicly funded Human Genome Project, launched in 1990. The whole-genome shotgun approach, championed by Craig Venter's company Celera, skipped the map and assembled directly from overlapping reads. The public project and Celera published draft human genomes in 2001; a finished reference followed in 2003. The competition pushed both sides to accelerate, and the public data-release norms of the consortium—daily deposition, free access—became a defining ethic of genomics.

Assembly software evolved alongside. The phred base-calling program (1998) converted raw fluorescence signals into letters with quality scores, and the phrap assembler used those scores. The shift to short reads in the mid-2000s (see below) forced new graph-based assemblers, which represent overlaps as a network and simplify it; the de Bruijn graph approach, borrowed from mathematics, became standard for short-read assembly.

The Sequencing Deluge and Reference-Based Analysis

Massively parallel sequencing, introduced commercially around 2005–2007, changed the economics: instead of one long read, machines produced millions of short ones. Sequencing a human genome fell from billions of dollars to thousands within a decade. The new bottleneck was analysis. Rather than assembling every genome from scratch, the field aligned short reads to a reference genome—a fast, approximate mapping problem solved by tools such as Bowtie (2009) and BWA (2009), which use an index of the reference to find candidate positions quickly. Once reads were placed, variant calling compared them to the reference to identify single-letter differences (SNPs) and small insertions or deletions (indels). The Genome Analysis Toolkit (GATK), developed at the Broad Institute from the late 2000s, became a widely used framework for this, with best-practice pipelines that handled local realignment and quality recalibration.

The same period produced the first widely used expression pipelines. RNA sequencing (RNA-seq), introduced around 2008, measures which genes are switched on by counting reads from transcribed RNA5. Tools such as TopHat and Cufflinks (2010–2012) aligned RNA reads across splice junctions and estimated transcript abundance; later, fast quantifiers like kallisto and Salmon (2015–2017) skipped full alignment and counted against a transcript index directly. These pipelines made functional genomics a computational discipline as much as a laboratory one.

Molecular Evolution and Phylogenetics

Sequence data made evolutionary trees computable. Distance-based methods (neighbor-joining, published by Naruya Saitou and Masatoshi Nei in 1987) cluster sequences by similarity; character-based methods such as maximum parsimony and maximum likelihood evaluate trees against the actual aligned columns. Maximum likelihood, implemented in programs like PHYLIP (Joseph Felsenstein, from the early 1980s) and later RAxML and MrBayes, became the workhorses of molecular phylogenetics. The molecular clock and neutral theory supplied the interpretive frame: differences accumulate at measurable rates, so trees can be dated and gene histories compared. Comparative genomics grew from the same soil—aligning whole genomes across species to find conserved regions, duplicated genes, and rearrangements.

Population and Statistical Genetics at Genomic Scale

Genome-wide data transformed population genetics. The International HapMap Project (2002–2010) catalogued common variation across human populations and showed that variants are correlated in blocks—linkage disequilibrium—which made it possible to genotype a subset of markers and infer the rest6. Genome-wide association studies (GWAS), beginning in earnest around 2005–2007, tested hundreds of thousands of markers against traits and diseases. The statistical machinery—mixed models to control for population structure, imputation to fill in ungenotyped variants, and later fine-mapping—became a core bioinformatics activity. Coalescent theory, developed by John Kingman in the early 1980s, provided the probabilistic model of how lineages merge backward in time, and it underpins modern methods for inferring demographic history from genomes.

Metagenomics: Reading Communities

Sequencing environmental samples without isolating organisms created a new classification problem. Metagenomics—the study of all the genetic material in a sample—began with environmental clone libraries in the 1990s and exploded with shotgun sequencing in the 2000s. The questions are different from single-genome work: which organisms are present, in what proportions, and what functions do their genes encode? Tools such as QIIME (2010) and mothur (2009) standardized the analysis of marker-gene surveys (especially the 16S ribosomal RNA gene for bacteria), while assemblers and binning algorithms attempted to reconstruct individual genomes from mixed reads. The field made clear that most microbial diversity had never been cultured, and that bioinformatics was the only route to seeing it.

Machine Learning and the Deep-Learning Turn

Statistical learning entered genomics early—hidden Markov models for gene finding in the 1990s, support vector machines for classification in the 2000s. The deep-learning wave arrived around 2015. Convolutional neural networks, which learn local patterns in sequence, were applied to predict regulatory elements, splice sites, and chromatin state from DNA alone. AlphaFold, DeepMind's protein-structure prediction system, announced in 2020, demonstrated that learned models could approach experimental accuracy on a problem that had resisted decades of effort. The methods are powerful but data-hungry and hard to interpret, and the field continues to debate when a learned model captures biology and when it captures the quirks of its training set.

Reproducibility, Workflows, and the Cloud

As pipelines grew into dozens of tools with version-specific behavior, reproducing a result became a problem in itself. Workflow systems—Snakemake (2012), Nextflow (2013), and the Common Workflow Language—let researchers describe analyses as explicit, rerunnable graphs. Containerization (Docker, Singularity) froze software environments. Cloud platforms and large public compute resources moved analysis off individual laptops. The norms of open data and open code, inherited from the Human Genome Project, became technical requirements: a result that cannot be rerun is increasingly treated as incomplete.

Pangenomes and Graph Representations

The single linear reference genome has a known flaw: it represents one individual's sequence and forces everyone else's variation into a comparison against it, which loses information in highly variable regions. Graph-based representations, which store variation as alternative paths through a network, address this. The idea has roots in earlier assembly graphs, but it matured into a practical approach in the late 2010s, and the Human Pangenome Reference Consortium, formed in 2019, began assembling reference genomes from many individuals. Pangenome graphs are now used for read alignment, variant calling, and comparative analysis, and they represent the field's current answer to the limits of a single reference.

Where the Field Stands

Today's bioinformatics is layered: public databases and archives underneath; alignment, assembly, and variant-calling tools in the middle; and statistical, evolutionary, and machine-learning analysis on top, increasingly orchestrated by reproducible workflow systems and run at cloud scale1. The open questions are the ones the history keeps returning to—how to represent variation without a single reference, how to make learned models interpretable, how to integrate data across scales from single cells to populations, and how to keep analyses reproducible as the software stack grows. Each generation of methods has answered a question the previous data could not, and each has left the next generation a larger and messier archive to read.

References

  1. Bioinformatics (Sequences, Sequence assembly, introduction, Goals), Wikipedia
  2. Needleman–Wunsch algorithm, Wikipedia
  3. Smith–Waterman algorithm (introduction, History), Wikipedia
  4. European Nucleotide Archive (introduction, History), Wikipedia
  5. RNA-Seq (introduction, History), Wikipedia
  6. International HapMap Project (introduction, Background), Wikipedia

The frameworks of Bioinformatics

Each framework is a way of working the field has used: when it held, and the question it answers.

  1. 1965-Present

    Sequence Databases and Curation

    How are biological sequences stored and curated?

    Fly to its tower in Nousfarer

  2. 1968-Present

    Molecular phylogenetics

    How are species related by their sequences?

    Fly to its tower in Nousfarer

  3. 1970-Present
  4. 1990-Present

    Sequence assembly

    How do we reconstruct a genome from fragments?

    Fly to its tower in Nousfarer

  5. 1995-Present

    Machine Learning and Deep Learning in Genomics

    How can machines learn patterns from genomic data?

    Fly to its tower in Nousfarer

  6. 1998-Present

    Metagenomics

    Who is present in an environmental sample?

    Fly to its tower in Nousfarer

  7. 2002-Present

    Statistical and Population Genomics

    How does genetic variation differ across populations?

    Fly to its tower in Nousfarer

  8. 2005-Present
  9. 2008-Present
  10. 2012-Present
  11. 2019-Present

    Pan-genome graph construction

    How do we represent all variation in a species?

    Fly to its tower in Nousfarer

How the ideas connect

Each arrow runs from an earlier idea to the later one it shaped. The list below says what passed between them.

Sequence Databases and CurationMolecular phylogeneticsPairwise Sequence AlignmentHeuristic Database SearchSequence assemblyMachine Learning and Deep Learning in…MetagenomicsStatistical and Population GenomicsReference-Based Read Mapping and Vari…Transcriptomic Quantification Pipelin…Reproducible Workflow and Cloud Compu…Pan-genome graph construction

Beyond the field

Where Bioinformatics's ideas came from in other fields, and what they led to there: the strongest 12 of 61 links.

A city in Nousfarer: a red-carpeted stair climbing terraces to a crowned tower on an acropolis, more towers round it, forest and snowy mountains beyond.
A city in Nousfarer, the game played in the browser.

Bioinformatics is a planet in Nousfarer

Its city has a tower for each of these 12 frameworks. Land, walk in, and climb one concept at a time.

Land on Bioinformatics

Related fields

Fields its ideas are linked to

More in Genomics

See where Bioinformatics sits on the Map of Human Knowledge