Appendix A — Glossary

Every term defined here is used somewhere in the book. The definitions are written to be read on their own, so a reader who meets an unfamiliar word in Chapter 11 does not have to reread Chapter 4 to make sense of it. Where a term only makes sense alongside another, the entry says so with see also. File formats are described by what they store rather than by their syntax; tools are described by the job they do rather than by their options.

A.1 A

Ab initio gene prediction: Identification of genes from sequence composition alone — codon usage, open reading frame length, ribosome binding site signals — without homology evidence from other organisms. Fast and fully automatic, but blind to very short genes and to the correct translation table unless told. See also Structural annotation, Prodigal.
Accession number: The stable identifier an archive assigns to a record, such as NM_000546.6 or P04637. The suffix after the full stop is a version: the accession names the record, the version names one exact sequence, and a reproducible analysis cites the version.
ACMG/AMP criteria: The framework published by the American College of Medical Genetics and Genomics with the Association for Molecular Pathology for classifying sequence variants into five tiers from benign to pathogenic by combining weighted evidence of several kinds. It is a procedure for weighing evidence, not a lookup table. See also Variant of uncertain significance.
Adapter: A short synthetic oligonucleotide ligated to a DNA fragment so that it can bind the sequencing flow cell. Adapter sequence read through at the end of a short insert must be removed before alignment. See also Trimming.
Alignment: An arrangement of two or more sequences, with gaps inserted, that places residues believed to descend from the same ancestral position in the same column. An alignment is a hypothesis produced by an algorithm, not an observation. See also Global alignment, Local alignment, Multiple sequence alignment.
Allele: One of the alternative sequences found at a genetic locus in a population.
Allele frequency: The proportion of chromosomes in a population carrying a given allele. Frequency is the single most used piece of evidence in clinical variant interpretation, and its power is unequal across populations because reference databases sample them unevenly. See also Minor allele frequency, gnomAD.
AlphaFold: A deep-learning method, and the associated database, that predicts protein three-dimensional structure from sequence with per-residue confidence estimates. A predicted model is a hypothesis with an attached uncertainty. See also pLDDT, Predicted aligned error.
Amino acid: The monomer unit of a protein; twenty are encoded by the standard genetic code. Their side-chain chemistry is what substitution matrices attempt to summarise.
Annotation: The assignment of coordinates and biological meaning to features in a sequence. Divided into structural annotation, which says where features are, and functional annotation, which says what their products do.
Assembly: The reconstruction of a genome sequence from overlapping reads, and also the resulting set of contigs or chromosomes. Assembly quality is described by contiguity, completeness and correctness, which can vary independently. See also Contig, BUSCO.
Asymmetric unit: The smallest part of a crystal from which the whole can be generated by symmetry, and what a PDB coordinate file actually contains. It need not correspond to the biological assembly, so three protein chains in a file may be one crystallographic arrangement rather than a trimer.
AUGUSTUS: A gene prediction program for eukaryotic genomes that models exon–intron structure with a generalised hidden Markov model, and can be guided by transcript or protein evidence.

A.2 B

BAM: The binary, compressed form of SAM. When coordinate-sorted and indexed it supports random access to any genomic region, which is why alignment files are distributed in this form. See also SAM, CRAM, Index.
Batch effect: Systematic technical variation associated with when or where samples were processed, which can exceed the biological effect of interest. Adjustable when batch and biological group are crossed; irrecoverable when they are confounded. See also Confounding.
Benjamini–Hochberg procedure: A step-up method that controls the false discovery rate by comparing ranked p-values against a proportionally increasing threshold. The default correction in genomics, where thousands of hypotheses are tested at once. See also False discovery rate.
B-factor: A column of a PDB coordinate file reporting positional uncertainty or thermal motion for each atom in an experimentally determined structure. In a predicted model the same column usually holds a confidence score instead, which must not be read as a temperature factor. See also pLDDT.
BGZF: Blocked GNU Zip Format: a gzip-compatible compression in which data are stored in independent blocks of at most 64 kb, each with a recorded virtual offset. This is what allows a compressed VCF or BAM to be indexed and read at a single genomic position. See also Tabix, Index.
Bioconda: A community-maintained distribution channel supplying bioinformatics software as conda packages, which is how most command-line tools in this book are installed and pinned. See also Conda.
Bioconductor: A coordinated collection of R packages for the analysis of high-throughput biological data, built around shared data structures so that packages interoperate. See also SummarizedExperiment.
Biopython: A Python library providing sequence, alignment, structure and database objects, including the SeqIO and PDB modules used throughout this book.
Biostrings: The Bioconductor class and package for biological sequences in R, providing memory-efficient storage and vectorised operations such as base composition, pattern matching and reverse complementation. See also Bioconductor.
Bit score: An alignment score normalised so that it is independent of the scoring system and of the database searched, which is why it is comparable between searches while an E-value is not. See also E-value.
BLAST: The Basic Local Alignment Search Tool, a heuristic method that finds local alignments between a query and a sequence database by extending short exact matches, and reports each hit with an E-value. Fast enough to search an archive, and sensitive enough for most homology questions. See also E-value, Local alignment.
BLOSUM: A family of amino acid substitution matrices estimated from aligned blocks of related proteins. The number names the identity level of the blocks used, so BLOSUM45 suits distant comparisons and BLOSUM80 close ones. See also Scoring matrix.
Broad consent: A consent model in which participants permit future health-related research in general, with specific uses overseen by an ethics committee and a data access committee rather than approved individually. It keeps cohorts usable for decades at the cost of making the participant’s agreement thin. See also Tiered consent, Dynamic consent.
BUSCO: A method and a set of benchmarking gene collections used to estimate how complete an assembly or annotation is, by counting how many near-universal single-copy orthologues are present, duplicated, fragmented or missing.
BWA: The Burrows–Wheeler Aligner, a short-read aligner that indexes the reference genome with a compressed suffix array; bwa mem is the standard mode for reads of 70 bp and longer. See also Mapping quality.

A.3 C

CDS: Coding sequence: the portion of a transcript that is translated, beginning at a start codon and ending at a stop codon. In GFF3 the CDS feature is a child of a transcript feature. See also Open reading frame.
Central dogma: The model of information flow in cells from DNA to RNA to protein, with replication maintaining the DNA copy. Each step explains why a particular representation exists in bioinformatics.
Centrality: Any measure of how important a node is in a network. Degree centrality counts a node’s edges and identifies hubs; betweenness centrality counts the shortest paths running through a node and identifies connectors between modules. See also Hub.
CIGAR: The field of a SAM record describing how a read aligns to the reference as a run-length string of operations — match or mismatch, insertion, deletion, soft or hard clipping. The CIGAR, not the read length, determines where an alignment ends. See also SAM, Soft clipping.
ClinVar: A public archive of submitted assertions about the clinical significance of human variants, with the evidence, the submitter and a review status recorded for each. The review status matters more than the significance label. See also ACMG/AMP criteria.
Clustering: Grouping observations by similarity without using labels. The result depends on the distance measure, the linkage or algorithm and the number of groups requested, none of which the data determine on their own. See also Unsupervised learning.
Coding density: The percentage of a genome covered by coding sequence. A useful annotation diagnostic: in compact bacterial genomes it approaches 95%, and a prediction far below the expected value has lost or fragmented genes.
Codon: A triplet of nucleotides specifying one amino acid or a translation stop.
Codon usage bias: The unequal use of synonymous codons within a genome, which differs between organisms and between highly and lowly expressed genes. Gene finders exploit it to distinguish coding from non-coding sequence.
Coercion: Automatic conversion of values to a common type, which in R promotes a mixed vector to the most general type present, so one character value turns a whole numeric vector into text. A numeric column that will not do arithmetic has almost always been coerced. See also Factor.
Community detection: Algorithms that partition a network into groups with more internal than external edges. They always return a partition, so the diagnostic question is whether the modularity is high and whether different algorithms agree. See also Modularity, Module.
Conda: A package and environment manager that installs software and its dependencies into isolated environments described by a specification file, which is how this book pins its software. See also Bioconda, Environment specification.
Confounding: The situation in which two possible explanations for a result cannot be separated because they vary together, as when every treated sample is sequenced in one run. No statistical adjustment repairs a confounded design. See also Batch effect.
Confusion matrix: The table of predicted against true classes from which accuracy, precision, recall and specificity are all computed. Reporting it avoids the ambiguity of any single summary figure. See also Precision, Sensitivity.
Consent: The permission under which data about a person may be collected and used. In genomics the practical question is not whether consent exists but whether it covers the analysis intended, and whether the restriction is enforced in the sample table and the pipeline.
Contact map: A matrix in which entry (i, j) is set when residues i and j of a structure lie closer than a chosen cut-off, typically 8 Å between Cα atoms. Bands running perpendicular to the diagonal indicate antiparallel β-strands.
Contig: A contiguous sequence assembled from overlapping reads. An assembly’s contiguity is usually summarised by the number of contigs and by N50, the length such that half the assembly lies in contigs at least that long.
Controlled access: A governance model in which individual-level data remain in a managed repository and investigators apply to a data access committee for permission to use them for a stated purpose. The standard arrangement for human genomic data, which cannot be anonymised. See also dbGaP, European Genome-phenome Archive, Five safes.
Counts per million: A simple library-size normalisation in which each count is divided by the sample’s total and multiplied by a million, abbreviated CPM. It corrects depth but not composition, which is why differential expression methods use more robust size factors. See also Median-of-ratios normalisation.
Coverage: The average number of reads overlapping a position, also called depth. Coverage determines the confidence of a variant call, and uneven coverage is a diagnostic of capture bias, repeats or contamination.
CRAM: A reference-based compressed alignment format that stores differences from the reference rather than full read sequences, giving smaller files than BAM at the cost of requiring the exact reference to decode.
Cross-validation: Repeatedly fitting a model on part of the data and evaluating it on the held out remainder, to estimate performance on unseen data. Valid only if every fitting step, including feature selection and normalisation, happens inside the training part. See also Overfitting.
Curated database: A resource whose records are reviewed, corrected and made non-redundant by expert or automated curation, such as UniProtKB/Swiss-Prot or RefSeq, as distinct from a primary archive that stores what was deposited. See also Primary archive.

A.4 D

Data access committee: The body that assesses applications to use controlled-access data, judging whether the proposed use falls within participants’ consent, whether the applicant’s institution accepts the obligations, and whether the data requested are the minimum needed.
Data frame: R’s table type: a list of equal-length vectors presented as columns, each with its own type. See also Tibble.
Data management plan: A document written before data exist that states what will be collected, where it will live, who may access it, how long it is kept, what is deposited at publication and who remains responsible afterwards.
Data sovereignty: The principle that data about a people, or collected within a jurisdiction, is governed under that community’s or country’s authority rather than that of whoever holds a copy. See also Helicopter research.
dbGaP: The database of Genotypes and Phenotypes at NCBI, which holds controlled access human study data with openly discoverable study-level metadata.
dbSNP: An NCBI archive of short genetic variation, which assigns the rs identifiers used to refer to variants across resources. An rs number asserts that a variant was observed, not that it is common or benign.
Degree: The number of edges attached to a node in a network. See also Centrality, Hub.
De-identification: The removal of direct identifiers from a dataset. It guarantees only that those fields are gone; it does not make genomic data anonymous, because the genotype is itself an identifier. See also Pseudonymisation.
DESeq2: An R/Bioconductor package for differential expression analysis of count data, which models counts with a negative binomial distribution and shrinks dispersion and fold-change estimates towards a fitted trend. See also edgeR, Differential expression.
Differential expression: The comparison of transcript or gene abundance between conditions, testing for each gene whether the difference exceeds what the measurement’s noise model allows. Requires normalisation, a replicated design and multiple-testing control.
Dimensionality reduction: Projection of high-dimensional data onto a few axes for visualisation or downstream modelling, as in principal component analysis, t-SNE and UMAP. Separation in such a projection is a hypothesis, not a test result.
Dispersion: In a negative binomial model of count data, the parameter describing how much more variable the counts are than a Poisson process would predict. With few replicates it is estimated poorly per gene, so DESeq2 and edgeR shrink per-gene estimates towards a fitted mean–dispersion trend. See also DESeq2.
DNA: Deoxyribonucleic acid, the double-stranded polymer of nucleotides that stores hereditary information.
Domain: A structurally or functionally autonomous unit of a protein, which often folds independently and recurs across proteins. Domain content is the usual basis for functional annotation by similarity. See also Pfam, InterPro.
dplyr: The tidyverse package providing the data-manipulation verbs filter, select, mutate, arrange, group_by and summarise, composed with a pipe. See also Pipe, Tidy data.
DSSP: The standard algorithm for assigning secondary structure to a set of coordinates from backbone hydrogen-bonding patterns. See also Secondary structure.
Dual-use research of concern: Research conducted for legitimate purposes that could be directly misapplied to cause harm, such as work that increases a pathogen’s transmissibility or that reconstructs a dangerous agent from published sequence.
Duplicate marking: Flagging read pairs with identical alignment start positions as likely products of the same original fragment, so that variant callers do not count one molecule many times.
Dynamic consent: A consent model in which participants hold an account through which they can see what has been done with their data and change their permissions at any time. The most responsive model and the most demanding to operate.

A.5 E

edgeR: An R/Bioconductor package for differential expression of count data using negative binomial models and empirical Bayes moderation of dispersion estimates. See also DESeq2.
Enrichment analysis: Testing whether a gene set is over-represented in a category — a Gene Ontology term, a pathway — relative to a defined background. The background must be the set of genes the experiment could have detected, not the whole genome. See also Gene Ontology, Multiple testing.
Ensembl: A genome annotation resource and browser for vertebrate and other genomes, providing gene models, comparative genomics and variation data through a browser, an API and bulk files.
Environment specification: A file listing the exact software packages, versions and channels an analysis requires, from which the environment can be recreated. A version range is not a specification.
European Genome-phenome Archive: A controlled-access archive, operated by EMBL-EBI and the Centre for Genomic Regulation, for human data whose consent does not permit open distribution. Abbreviated EGA. See also Controlled access.
E-value: In a database search, the number of alignments with at least the observed score expected to occur by chance given the database size. Because it accounts for length and search space, it is a better guide to homology than percentage identity. See also BLAST.
Exome: The protein-coding portion of a genome, about 1–2% of human sequence, and by extension the capture-based sequencing strategy targeting it. See also Whole-exome sequencing.
Exon: A segment of a transcript retained after splicing. Exon boundaries are why a transcript’s coordinates differ from its gene’s genomic coordinates.

A.6 F

Factor: R’s type for a categorical variable, stored as integer codes plus a table of permitted levels. Because the codes are integers, arithmetic on a factor silently uses them, and only as.numeric(as.character(x)) recovers the labels.
FAIR principles: A set of expectations that data be findable, accessible, interoperable and reusable, through persistent identifiers, documented access procedures, standard formats and vocabularies, and clear licensing. FAIR does not mean open: controlled-access data can be fully FAIR.
False discovery rate: The expected proportion of false positives among the tests declared significant. Controlling it, rather than the probability of any false positive, is what makes genome-wide testing tractable. See also Benjamini–Hochberg procedure, Multiple testing.
FASTA: A plain-text format in which each record is a header line beginning with > followed by sequence lines. It carries sequence and a free-text description and nothing else — no quality, no coordinates, no annotation.
FASTQ: The read format that pairs each sequence with per-base quality scores, in four lines per record. The quality line is what distinguishes it from FASTA and what quality control operates on. See also Phred quality score.
FastQC: A program that computes per-base and per-read quality summaries for sequencing data and presents them as a report. Its red and amber flags are heuristics tuned for whole-genome data and must be read with the library type in mind.
featureCounts: A program that assigns aligned reads to genomic features such as exons or genes, producing the count matrix used in differential expression analysis.
Five safes: A framework that decomposes safe data access into five separable questions — safe projects, safe people, safe settings, safe data and safe outputs — so that weakness in one dimension can be compensated in another.
FLAG: The second field of a SAM record: a bit field encoding whether the read is paired, properly paired, mapped, reverse-stranded, the first or second mate, a duplicate, or a secondary or supplementary alignment. See also SAM, Proper pair.
flagstat: The samtools subcommand that counts the alignment records in a BAM file by FLAG category — total, mapped, properly paired, singletons, duplicates — and the usual first summary of an alignment. See also SAMtools, FLAG.
FM-index: A compressed full-text index built on the Burrows–Wheeler transform that supports exact substring search in space close to that of the compressed text. It is what allows a short-read aligner to index a whole genome and place millions of reads against it. See also BWA.
Frameshift: An insertion or deletion whose length is not a multiple of three, which changes the reading frame of every downstream codon and usually introduces a premature stop. Generally more damaging than a single missense change.
Functional annotation: Assignment of biological meaning — product name, domains, Gene Ontology terms, pathway membership — to a predicted feature, almost always by transfer from homologous sequences. See also Structural annotation, Homology.

A.7 G

GATK: The Genome Analysis Toolkit, a suite for variant discovery from high-throughput sequencing data, including the local reassembly-based HaplotypeCaller and the filtering steps that follow it.
GC content: The percentage of bases in a sequence that are G or C. It varies systematically between genomes and along them, affects sequencing coverage and gene-prediction models, and is the quickest compositional check for contamination.
GenBank: The NCBI primary nucleotide sequence archive, and the flat-file format in which its records are distributed, carrying sequence together with a feature table. Part of the International Nucleotide Sequence Database Collaboration with ENA and DDBJ. See also Primary archive.
Gene: A segment of DNA transcribed as a unit, together with the regulatory sequence governing it. The operational definition varies between resources, which is why gene counts differ between annotations of the same genome.
Gene Ontology: A structured, species-independent vocabulary describing gene product molecular functions, biological processes and cellular components, with terms arranged in a hierarchy and annotations carrying evidence codes.
Gene set enrichment analysis: A method, abbreviated GSEA, that tests whether members of a gene set are concentrated at one end of a ranked list of all genes, rather than testing a thresholded subset. It detects coordinated small shifts that over-representation analysis misses, at the cost of depending on the ranking statistic. See also Enrichment analysis.
Genetic code: The mapping from codons to amino acids and stop signals. Variants of the standard code exist: NCBI translation table 4, used by Mycoplasma and relatives, reads TGA as tryptophan rather than as a stop, which breaks gene prediction run under the bacterial default. See also Translation table.
Genome browser: An interactive viewer that displays annotation, alignment and variation tracks against genomic coordinates, such as Ensembl, the UCSC browser or IGV. Looking at the reads is the fastest way to reject a spurious variant call.
Genome-wide association study: A study that tests association between many common variants and a phenotype across a population, correcting for multiple testing and population structure. Abbreviated GWAS.
GFF3: A nine-column tab-separated annotation format giving each feature’s sequence, type, coordinates, strand, phase and attributes, with ID and Parent attributes expressing the gene–transcript–CDS hierarchy. Coordinates are one-based and inclusive. See also GTF.
ggplot2: The R implementation of the grammar of graphics, in which a plot is built from data, aesthetic mappings, geometries, scales, facets and a theme, added as layers. It requires data in long form. See also Tidy data.
Git: A distributed version control system that records the history of a set of files as a series of commits. In an analysis project it holds code, configuration and documentation, never derived results or controlled-access data. See also Version control.
Global alignment: An alignment of two sequences along their entire length, appropriate when they are expected to correspond end to end. Computed optimally by the Needleman–Wunsch algorithm.
gnomAD: A reference database of allele frequencies aggregated from large exome and genome collections, used as the population-frequency evidence in variant interpretation. Its ancestry composition determines how much power that evidence has for a given patient.
GRanges: The Bioconductor class for genomic intervals, each carrying a sequence name, start, end and strand plus arbitrary metadata columns, with set operations and overlap queries defined on it. See also Bioconductor.
GTF: A tab-separated annotation format closely related to GFF, with a fixed attribute convention centred on gene_id and transcript_id. Widely expected by RNA sequencing tools. See also GFF3.

A.8 H

Haplotype: A combination of alleles at linked sites inherited together on the same chromosome. See also Linkage disequilibrium.
Hardy–Weinberg equilibrium: The expected relationship between allele and genotype frequencies in a large randomly mating population, which gives genotype frequencies $p^2$, $2pq$ and $q^2$. Departures are used as a quality filter on genotype data and as evidence of population structure.
Helicopter research: Research in which investigators from a well-resourced institution collect samples or data in a lower-income setting, analyse them elsewhere, and leave no capacity, authorship or benefit behind. Also called parachute or extractive research. See also Data sovereignty.
Hidden Markov model: A probabilistic model of a sequence of observations generated by unobserved states with defined transition probabilities. Used for gene structure prediction, protein family profiles and secondary structure assignment. See also HMMER, AUGUSTUS.
HIPAA: The US Health Insurance Portability and Accountability Act, which regulates protected health information held by covered entities such as providers and health plans and by their business associates. Many research laboratories are not covered entities, and genomic sequence is not among the identifier types its Safe Harbor standard enumerates.
HMMER: A program suite that searches sequences against profile hidden Markov models and builds such profiles from alignments, giving greater sensitivity to distant homology than pairwise search. See also Pfam, Hidden Markov model.
Homology: Shared ancestry between sequences or structures. Homology is inferred from similarity, never measured directly, and sequences are either homologous or not — “percentage homology” is a category error. See also Orthologue, Paralogue.
Homology modelling: Prediction of a protein structure by aligning its sequence to that of an experimentally determined relative and copying the template’s coordinates. Its accuracy is governed mainly by sequence identity to the template, and it cannot predict a fold absent from the database.
Hub: A network node with unusually high degree. Hubs are often essential, but in literature-derived networks they are also the best-studied proteins, which is a confound rather than a finding. See also Degree.
Hypergeometric test: The exact test for over-representation of a category in a sample drawn without replacement, and the statistic underneath most enrichment tools: it asks how improbable the observed overlap between a gene list and a term would be if the list were drawn at random from the background universe. See also Enrichment analysis.

A.9 I

Incidental finding: A result of potential clinical significance discovered outside the question a study set out to answer. Whether to return one should be decided in the protocol, before sequencing, and by the people with clinical and ethical standing to decide. See also Secondary finding.
Indel: An insertion or deletion of one or more bases, reported as a single variant class because alignment often cannot distinguish which event occurred. See also Frameshift.
Index: An auxiliary file that maps coordinates or identifiers to byte offsets in a large data file so that a region can be retrieved without reading the whole file. Required for BAM, CRAM, compressed VCF and indexed FASTA. See also Tabix.
Informed consent: The process by which a participant, having been given an adequate explanation of a study’s purposes, procedures, risks and data uses, voluntarily agrees to take part. See also Broad consent.
Insert size: The length of the original DNA fragment from which a read pair was sequenced, reported in SAM as the signed template length. Its distribution is a property of library preparation, and an implausibly tight one is a sign of simulated data.
InterPro: An integrative resource that combines protein family, domain and site signatures from several member databases, and the InterProScan tool that searches a sequence against all of them at once. See also Pfam, Domain.
Intron: A segment of a primary transcript removed by splicing. Introns are why eukaryotic gene prediction is substantially harder than bacterial gene prediction.
Isoform: One of several distinct transcripts or protein products of the same gene, typically arising from alternative splicing or alternative promoters. A protein record’s canonical isoform may differ from the transcript you translated.

A.10 K

KEGG: A collection of manually drawn pathway maps linking genes and metabolites to molecular interaction and reaction networks, widely used as the pathway vocabulary for enrichment analysis.
k-means clustering: An algorithm that partitions observations into a pre-specified number of clusters by iteratively assigning each to the nearest centroid and recomputing centroids. The number of clusters is an input, not a result. See also Clustering.
k-mer: A substring of length k. Counting k-mers underlies assembly graph construction, sequence comparison without alignment, and contamination screening.

A.11 L

Leakage: Any route by which information from held-out data influences model fitting — most often feature selection, normalisation or hyperparameter tuning performed before the data were split. It inflates reported performance, and label permutation is the standard test for it. See also Cross-validation.
Licence: The terms under which others may use a work. Code, text and data follow different conventions — permissive and copyleft software licences, Creative Commons for text and figures, open data licences for databases — and unlicensed code is by default not reusable.
Linkage disequilibrium: The non-random association of alleles at nearby loci, arising because recombination has not yet separated them. It makes genome-wide association possible, by letting a typed marker tag an untyped causal variant, and it is also why polygenic scores transfer poorly between populations with different haplotype structure. See also Polygenic score.
Local alignment: An alignment of the best-scoring subsequences of two sequences, appropriate when only part of each is expected to correspond. Computed optimally by the Smith–Waterman algorithm and heuristically by BLAST.
Long-read sequencing: Sequencing technologies producing reads of thousands to millions of bases, at a higher per-base error rate than short reads, which resolve repeats and structural variation that short reads cannot.

A.12 M

MAFFT: A program for multiple sequence alignment offering a range of algorithms from fast progressive methods for large inputs to accurate iterative refinement for small ones.
Makefile: A file declaring, for each output, the inputs and the command that produces it, from which make computes what is out of date and rebuilds only that. It records dependencies rather than execution order. See also Workflow manager.
MA plot: A scatter plot of log fold change against mean expression used to inspect a differential expression result. A systematic vertical shift indicates a normalisation problem rather than biology.
Mapping quality: A read’s alignment confidence, expressed on the Phred scale, which reports the probability that the read came from somewhere else in the reference. Low values indicate repeats or paralogues and are a standard variant filter.
MAPQ: The mapping quality field of a SAM record, a Phred-scaled estimate of the probability that the read was placed in the wrong location. A MAPQ of 0 means the read aligns equally well elsewhere, usually in a repeat, and says nothing about read quality. See also Mapping quality, SAM.
Median-of-ratios normalisation: The size-factor estimator used by DESeq2: each sample’s factor is the median across genes of its count divided by the geometric mean of that gene across samples. It is robust to a minority of strongly changing genes in a way that total-count scaling is not. See also Counts per million, DESeq2.
Minor allele frequency: The frequency of the less common allele at a site, abbreviated MAF. Rare variants carry little individual-identifying information whereas common ones carry a great deal, which is why forensic and sample-tracking panels use common markers. See also Allele frequency.
Missense variant: A single-nucleotide change that alters the encoded amino acid. Its consequence depends on the residue’s structural and functional context, which is why prediction of missense effect remains imprecise.
mmCIF: The macromolecular Crystallographic Information File format, the current standard representation for structures in the Protein Data Bank, which unlike the legacy PDB format is not limited by fixed column widths. See also PDB.
Modularity: A score, written Q, measuring how many more edges fall inside a proposed partition’s groups than would be expected at random. Low modularity across several algorithms indicates that the communities returned are artefacts of the method. See also Community detection.
Module: A densely interconnected group of nodes in a network, often corresponding to a protein complex or a pathway. See also Clustering.
Multiple sequence alignment: An alignment of three or more sequences, abbreviated MSA, from which conserved columns, phylogenies and profile models are derived. Progressive methods build it from pairwise alignments, so an early error propagates. See also MAFFT.
Multiple testing: The inflation of false positives that occurs when many hypotheses are tested at once, and the corrections that control it. A per-test p-value threshold of 0.05 across 20,000 genes expects 1,000 false positives. See also False discovery rate.
Mutation: A change in a DNA sequence relative to a reference or a parental sequence. In clinical contexts “variant” is now preferred, because “mutation” carries an implication of pathogenicity that the observation does not support.
Mutational signature: The characteristic distribution over mutation types, conventionally 96 classes of substitution in trinucleotide context, that a particular mutational process generates. A tumour’s spectrum is modelled as a mixture of signatures; the result describes chemistry, not functional consequence.

A.13 N

N50: The length such that half of an assembly lies in contigs at least that long. It measures contiguity rather than correctness and rewards over-aggressive assembly, so it should be reported with total length, contig count and a completeness estimate. See also Contig, BUSCO.
Nagoya Protocol: An international agreement under the Convention on Biological Diversity governing access to genetic resources and the fair and equitable sharing of benefits from their use, implemented through prior informed consent from the provider country and mutually agreed terms. Its application to digital sequence information remains under negotiation.
NCBI: The US National Center for Biotechnology Information, which operates GenBank, RefSeq, PubMed, dbSNP, ClinVar, the Sequence Read Archive and the BLAST services.
NDPR: The Nigeria Data Protection Regulation of 2019, which established consent requirements, data subject rights and cross-border transfer restrictions for personal data in Nigeria; the Nigeria Data Protection Act 2023 placed the regime on a statutory footing with a dedicated commission as regulator.
Needleman–Wunsch algorithm: The dynamic programming algorithm that computes an optimal global alignment of two sequences under a given scoring scheme. See also Global alignment.
Next-generation sequencing: The massively parallel sequencing technologies that replaced capillary sequencing, producing millions of short reads per run. Abbreviated NGS.
Nonsense-mediated decay: A surveillance pathway that degrades transcripts bearing a premature termination codon, which is why many nonsense and frameshift variants reduce protein to near zero rather than producing a truncated product.
Nonsense variant: A single-nucleotide change that converts an amino acid codon into a stop codon, truncating the protein. See also Nonsense-mediated decay.
Normalisation: Adjustment of measurements so that they are comparable across samples, for instance correcting sequencing counts for differences in library size and composition. A raw count is not an abundance.
Nucleotide: The monomer unit of DNA and RNA, comprising a base, a sugar and a phosphate.

A.14 O

Open reading frame: A stretch of sequence in one reading frame that contains no stop codon, conventionally from a start codon to the next in-frame stop. An ORF is a candidate gene, not a gene. See also Ab initio gene prediction.
Orthologue: A homologous sequence in another species that descends from the same ancestral sequence by speciation, and which usually retains the ancestral function. See also Paralogue.
Overfitting: Fitting the noise in the training data rather than the signal, so that performance on new data is much worse than on the data used to build the model. The main risk when features greatly outnumber samples. See also Cross-validation.

A.15 P

Paired-end sequencing: Sequencing both ends of each fragment, giving two reads of known relative orientation and approximate separation, which improves alignment in repetitive regions and allows detection of structural variation.
Paralogue: A homologous sequence within the same genome that arose by duplication. Paralogues may have diverged in function, so transferring annotation between them is less safe than between orthologues.
Pathogenicity: The degree to which a variant causes disease. Classified from weighed evidence rather than measured, and revised as evidence accumulates. See also ACMG/AMP criteria.
PDB: The Protein Data Bank, the archive of experimentally determined macromolecular structures, and the legacy fixed-column coordinate file format it introduced. See also mmCIF, B-factor.
Permutation test: A test in which the labels or the data are shuffled many times to build the null distribution of a statistic empirically. In machine learning it is the standard check for leakage: with labels permuted, an honest pipeline must perform no better than chance. See also Leakage.
Pfam: A database of protein families represented as profile hidden Markov models built from curated seed alignments, now distributed as part of InterPro. See also HMMER.
Pharmacogenomics: The study of how inherited variation affects drug response, and the clinical application of that knowledge through gene–drug guidelines covering dosing and drug choice.
Phred quality score: A base-calling confidence expressed as $-10\log_{10}$ of the error probability, so that 20 means one error in 100 and 30 means one in 1,000. Encoded as ASCII characters in FASTQ. See also FASTQ.
Phylogenetics: The inference of evolutionary relationships from molecular data, usually presented as a tree whose topology, branch lengths and support values each carry different information.
Pileup: A per-position view of the bases and qualities of all reads covering a site, and the input to most variant callers. Inspecting the pileup is the fastest way to reject an artefactual call.
Pipe: An operator that passes the output of one expression as the input of the next: | in the Unix shell, |> in base R since version 4.1 and %>% in the magrittr and tidyverse idiom. It makes a sequence of transformations readable in the order they happen.
Pipeline: An ordered set of analysis steps, each consuming the previous step’s output. Written as a workflow definition rather than as a list of commands, it becomes reproducible and restartable.
pLDDT: The per-residue confidence score of an AlphaFold model, on a 0–100 scale, stored in the B-factor column. Low values often indicate genuine disorder rather than prediction failure. See also AlphaFold.
Polygenic score: A weighted sum of many variant dosages used to estimate an individual’s genetic liability to a trait. Weights estimated in one population transfer poorly to others because the markers tag causal variants through population-specific linkage disequilibrium.
Precision: The proportion of positive predictions that are correct. Reported alongside sensitivity, because either alone can be made perfect by a trivial predictor. See also Sensitivity.
Precision–recall curve: A plot of precision against recall across classification thresholds, and the appropriate summary when the positive class is rare, because unlike a receiver operating characteristic curve it is not flattered by a large number of true negatives. See also ROC curve, Precision.
Predicted aligned error: A matrix accompanying a predicted structure giving the expected positional error of residue i when the model is superposed on residue j. It is the quantity to consult before trusting the relative placement of two domains. See also AlphaFold.
Primary archive: A repository that stores records as deposited, with the submitter responsible for their content and earlier versions retained, such as GenBank or the Sequence Read Archive. See also Curated database.
Principal component analysis: A linear projection onto orthogonal axes ordered by the variance they capture, used to visualise structure and to summarise correlated features. The components are directions of variance, not of biological interest. See also Dimensionality reduction.
Prodigal: An ab initio gene prediction program for prokaryotic genomes, which trains a coding model on the genome itself and scores open reading frames by codon composition and ribosome binding site signals. Its default translation table is the bacterial standard, which is wrong for some clades. See also Translation table.
Prokka: A pipeline that wraps gene prediction with RNA gene finding and functional annotation by homology to produce a complete prokaryotic annotation in standard formats.
Proper pair: A read pair that the aligner considers correctly placed: both mates mapped, in the expected relative orientation, and separated by a plausible insert size. The proportion of properly paired reads is a standard alignment diagnostic. See also flagstat.
Pseudogene: A sequence recognisably derived from a gene but no longer producing a functional product, through truncation, frameshift or loss of expression. Pseudogene loci are where annotation methods most often disagree.
Pseudonymisation: Replacement of direct identifiers with a code while retaining the ability to re-identify through a separately held key. Pseudonymised data remain personal data in law. See also De-identification.
p-value: The probability of observing data at least as extreme as those obtained if the null hypothesis were true. It is not the probability that the hypothesis is false, and it says nothing about effect size.

A.16 Q

Quality control: The checks applied to data before analysis to establish that they are fit for the question, and the diagnostics that detect contamination, bias or technical failure. The step whose omission costs the most time later.
QUAST: A tool that computes assembly quality metrics — contiguity, length distribution, misassemblies against a reference — and reports them in a single comparison table.

A.17 R

R: A language and environment for statistical computing and graphics, the usual choice in this book for modelling and figures, extended for biology by Bioconductor.
Radius of gyration: The root-mean-square distance of a structure’s atoms from its centre of mass, a single-number measure of compactness. For a model whose domains are joined by disordered linkers it describes one arbitrary arrangement rather than a property of the protein.
Ramachandran plot: A scatter plot of the backbone dihedral angles φ and ψ for every residue in a structure. Residues cluster in the sterically allowed α-helical and β regions, and a point in a forbidden area is a modelling error until shown otherwise.
Random forest: An ensemble classifier or regressor built from many decision trees trained on bootstrap samples with random feature subsets, whose predictions are aggregated. Robust on tabular data and resistant, though not immune, to overfitting.
Read: A single sequence produced by a sequencing instrument, with per-base quality scores. A read is an observation; where it came from is an inference.
Recycling: R’s rule of repeating the shorter operand of a vectorised operation until the lengths match, without warning when they divide evenly. On a matrix it recycles down the columns, so dividing a gene-by-sample count matrix by a vector of library sizes scales the wrong dimension and returns a plausible, wrong answer.
Reference genome: An assembled, coordinate-stable sequence used as the frame against which reads are aligned and variants reported. Results are only comparable between analyses using the same build, which is why the build identifier belongs in every methods section.
RefSeq: The NCBI curated, non-redundant collection of genomic, transcript and protein sequences, whose identifiers carry prefixes such as NM_ for mRNA and NP_ for protein. See also Curated database.
Reproducibility: The property that another person, given the same data, can obtain the same numbers. In practice it requires code under version control, a pinned environment, recorded parameters and seeds, and documented data provenance.
Resolution: For a crystal structure, the smallest spacing of diffraction data used, in ångströms, which bounds the detail the model can support: individual atoms near 1.5 Å, backbone and large side chains only beyond 3.5 Å. See also R factor.
R factor: The disagreement between the diffraction amplitudes observed and those calculated from a crystallographic model. R-free is the same quantity computed on reflections withheld from refinement, and is the meaningful one because it cannot be improved by overfitting.
Ribosome binding site: A sequence motif upstream of a start codon, in bacteria the Shine–Dalgarno sequence, which positions the ribosome. Gene finders model it to choose between candidate start codons.
RMSD: Root-mean-square deviation: the square root of the mean squared distance between corresponding atoms of two superposed structures. Meaningless without the number of atoms and the selection used. See also Superposition.
RNA: Ribonucleic acid, the single-stranded nucleotide polymer that carries genetic information from DNA to the ribosome and performs catalytic and regulatory roles.
RNA sequencing: Measurement of transcript abundance by sequencing cDNA derived from RNA, abbreviated RNA-seq, quantified either by counting reads over features or by assigning them to transcripts probabilistically. See also Differential expression.
ROC curve: A plot of true positive rate against false positive rate across classification thresholds, summarised by the area underneath it. On a strongly imbalanced problem it overstates performance. See also Precision–recall curve.

A.18 S

SAM: The Sequence Alignment/Map format, a tab-separated text representation of read alignments carrying position, mapping quality, a CIGAR string describing the alignment, mate information and optional tags. See also BAM, CRAM.
SAMtools: A program suite for manipulating alignment files — converting, sorting, indexing, filtering and computing pileups — and the companion to BCFtools for variant data.
Scoring matrix: A table of log-odds scores for aligning each pair of residues, encoding how likely that substitution is between homologues relative to chance. See also BLOSUM.
Secondary finding: A result of clinical significance that a project deliberately looks for beyond its primary question, as distinct from an incidental finding encountered unsought. See also Incidental finding.
Secondary structure: Local, regularly hydrogen-bonded elements of a polypeptide chain, chiefly α-helices and β-strands, assigned from coordinates by algorithms such as DSSP.
Sensitivity: The proportion of true positives that are detected, also called recall. In the annotation comparison of this book it is the proportion of reference coding sequences a prediction recovers. See also Precision.
Sequence Read Archive: The NCBI primary archive of raw sequencing reads, abbreviated SRA, mirrored by the European Nucleotide Archive.
Shine–Dalgarno sequence: A purine-rich ribosome binding motif a few bases upstream of a bacterial start codon, complementary to the 3’ end of the 16S ribosomal RNA. Gene finders model it to choose between candidate start codons. See also Ribosome binding site.
Short-read sequencing: Sequencing producing reads of roughly 50–300 bases at low per-base error rate, the basis of most resequencing, exome and RNA sequencing work.
Smith–Waterman algorithm: The dynamic programming algorithm that computes an optimal local alignment of two sequences. See also Local alignment.
Snakemake: A Python-based workflow manager that defines rules with input and output patterns, resolves the dependency graph, and executes it locally or on a cluster with environment isolation. See also Makefile, Workflow manager.
SNP: A single-nucleotide polymorphism: a single-base difference between chromosomes that is common in a population. A single-base difference without the frequency claim is better called a single-nucleotide variant.
Soft clipping: A CIGAR operation marking bases at the end of a read that are present in the record but not aligned, usually adapter sequence or accumulated errors. Hard-clipped bases, by contrast, are removed from the record entirely. See also CIGAR.
Specificity: The proportion of true negatives correctly identified. Required alongside sensitivity whenever classes are unbalanced.
Splicing: Removal of introns from a primary transcript and joining of exons, which is why transcript and genomic coordinates differ and why alignment of RNA reads to a genome must be splice-aware.
Standard streams: The three channels every Unix program has: standard input, standard output and standard error. Keeping results on standard output and messages on standard error is what allows programs to be joined by pipes without corrupting the data. See also Pipe.
STRING: A database of known and predicted protein–protein functional associations, integrating evidence channels into a confidence score. An edge asserts functional association, which is weaker than a physical interaction.
Structural annotation: Determination of where features lie in an assembly — genes, exons, RNA genes, repeats — as distinct from what their products do. See also Functional annotation.
SummarizedExperiment: The Bioconductor container holding one or more assay matrices together with sample and feature annotation, guaranteeing that subsetting cannot desynchronise them. See also Bioconductor.
Superposition: Rigid-body rotation and translation of one structure onto another to minimise the distance between corresponding atoms. The choice of atoms superposed determines what the resulting deviations mean. See also RMSD.
Supervised learning: Fitting a model to predict a known label or value from features, and evaluating it on data withheld from every stage of fitting. See also Unsupervised learning, Cross-validation.
Synonymous substitution: A nucleotide change that leaves the encoded amino acid unaltered, possible because the genetic code is degenerate. Synonymous changes are not necessarily neutral: they can affect splicing, stability and translation rate.

A.19 T

Tabix: An indexing method for position-sorted, block-compressed tab-delimited files such as VCF and GFF, which enables retrieval of a genomic interval without reading the file. See also Index.
Tibble: The tidyverse’s data frame, differing in three safety-relevant behaviours: no partial matching of column names, no dropping to a vector on single-column selection, and no silent conversion of character columns to factors. See also Data frame.
Tidy data: A table in which each variable is a column, each observation a row and each cell a value. It is the shape ggplot2 and grouped summaries require, and the opposite of the wide matrix that Bioconductor assays use; pivot_longer and pivot_wider convert between them.
Tiered consent: A consent model in which participants decide separately about distinct uses — commercial research, international deposition, re-contact, return of results. It respects differing preferences and fragments the cohort into subsets whose permissions must be enforced in code.
Transcription: Synthesis of RNA from a DNA template by RNA polymerase.
Transcriptome: The complete set of transcripts in a cell or sample, and their abundances, at a given time.
Transfer RNA: The adaptor RNA that carries an amino acid and recognises a codon through its anticodon, abbreviated tRNA. Predicted in genome annotation by specialised tools rather than by protein gene finders.
Translation: Synthesis of a polypeptide from a messenger RNA template by the ribosome, reading the sequence in triplets.
Translation table: One of the numbered NCBI genetic codes specifying which codons encode which amino acids and which terminate translation. Table 11 is the bacterial standard; table 4 applies to Mycoplasma and relatives; the table must be set before gene prediction. See also Genetic code.
Trimming: Removal of adapter sequence and low-quality bases from reads before alignment. Over-trimming discards real signal, so trimming parameters belong in the methods section. See also Adapter.
t-SNE: t-distributed stochastic neighbour embedding, a non-linear projection that preserves local neighbourhood structure for visualisation. Distances between clusters in a t-SNE plot carry no reliable meaning. See also UMAP, Dimensionality reduction.
Twilight zone: The band of roughly 20–35% sequence identity over a protein-length alignment in which real homologues and chance similarities both occur, so identity alone cannot separate them and a significance estimate is required. See also E-value.

A.20 U

UMAP: Uniform manifold approximation and projection, a non-linear dimensionality reduction method that preserves more global structure than t-SNE and runs faster, but whose output still depends strongly on its parameters.
UniProtKB: The central protein sequence and annotation resource, divided into the manually reviewed Swiss-Prot section and the automatically annotated TrEMBL section, with every annotation carrying an evidence attribution.
Unix: The family of operating systems, including Linux and macOS, whose shell and small composable text-processing programs form the working environment of bioinformatics.
Unsupervised learning: Finding structure in data without labels, as in clustering and dimensionality reduction. Its output is a hypothesis about structure, not a test of one. See also Supervised learning.
Untranslated region: The portion of a mature transcript upstream or downstream of the coding sequence, abbreviated UTR, which carries regulatory elements affecting stability and translation.

A.21 V

Variance-stabilising transformation: A transformation of count data that makes variance approximately independent of the mean, so that distance-based methods such as clustering and principal component analysis are not dominated by the most highly expressed genes.
Variant: A difference between an observed sequence and a reference at a defined position. The term is deliberately neutral about consequence. See also Mutation.
Variant calling: Inference of genotypes at positions where reads differ from the reference, followed by filtering to remove systematic artefacts. A call is an inference about a genotype, with a confidence, not a measurement.
Variant of uncertain significance: A variant whose available evidence is insufficient to classify as pathogenic or benign, abbreviated VUS. The classification describes the state of knowledge, which is why reclassification is routine and why uncertain results are commoner for under-represented populations. See also ACMG/AMP criteria.
VCF: The Variant Call Format, a tab-separated text format with a header of metadata definitions and one line per variant site, carrying position, alleles, quality, filter status, an extensible INFO field and per-sample genotypes. Compressed and indexed for random access. See also Tabix.
Vectorisation: Applying an operation to a whole array at once rather than element by element in a loop, which is both faster and less error-prone because no index is written by hand. It is the normal style in both NumPy and R.
Version control: Systematic recording of changes to files over time so that any previous state can be recovered and every change attributed. See also Git.
Volcano plot: A scatter plot of statistical significance against effect size for every gene in a differential expression analysis, used to show that the genes reaching significance also change by a worthwhile amount.

A.22 W

Whole-exome sequencing: Sequencing of the captured protein-coding portion of a genome, abbreviated WES, which concentrates coverage where interpretable variants are most dense at the cost of missing regulatory and structural variation.
Whole-genome sequencing: Sequencing of the entire genome without target capture, abbreviated WGS, giving more uniform coverage and access to non-coding and structural variation at greater cost per sample.
Workflow manager: Software that executes an analysis from a declarative description of its steps and their dependencies, rerunning only what is out of date and recording what was run. See also Makefile, Snakemake.

A.23 Y

Y-STR: A short tandem repeat marker on the Y chromosome. Because Y haplotypes and patrilineal surnames are inherited together, Y-STR profiles recovered from sequence data can be matched against genealogy databases to infer a surname.

A.24 Z

Z-score: A value expressed as the number of standard deviations from the mean of its distribution. Used to standardise features before clustering and to put measurements on comparable scales, and meaningful only if the reference distribution is stated.

--- title: "Glossary" --- Every term defined here is used somewhere in the book. The definitions are written to be read on their own, so a reader who meets an unfamiliar word in Chapter 11 does not have to reread Chapter 4 to make sense of it. Where a term only makes sense alongside another, the entry says so with *see also*. File formats are described by what they store rather than by their syntax; tools are described by the job they do rather than by their options. ## A {#sec-glossary-a} **Ab initio gene prediction** : Identification of genes from sequence composition alone — codon usage, open reading frame length, ribosome binding site signals — without homology evidence from other organisms. Fast and fully automatic, but blind to very short genes and to the correct translation table unless told. *See also* **Structural annotation**, **Prodigal**. **Accession number** : The stable identifier an archive assigns to a record, such as `NM_000546.6` or `P04637`. The suffix after the full stop is a version: the accession names the record, the version names one exact sequence, and a reproducible analysis cites the version. **ACMG/AMP criteria** : The framework published by the American College of Medical Genetics and Genomics with the Association for Molecular Pathology for classifying sequence variants into five tiers from benign to pathogenic by combining weighted evidence of several kinds. It is a procedure for weighing evidence, not a lookup table. *See also* **Variant of uncertain significance**. **Adapter** : A short synthetic oligonucleotide ligated to a DNA fragment so that it can bind the sequencing flow cell. Adapter sequence read through at the end of a short insert must be removed before alignment. *See also* **Trimming**. **Alignment** : An arrangement of two or more sequences, with gaps inserted, that places residues believed to descend from the same ancestral position in the same column. An alignment is a hypothesis produced by an algorithm, not an observation. *See also* **Global alignment**, **Local alignment**, **Multiple sequence alignment**. **Allele** : One of the alternative sequences found at a genetic locus in a population. **Allele frequency** : The proportion of chromosomes in a population carrying a given allele. Frequency is the single most used piece of evidence in clinical variant interpretation, and its power is unequal across populations because reference databases sample them unevenly. *See also* **Minor allele frequency**, **gnomAD**. **AlphaFold** : A deep-learning method, and the associated database, that predicts protein three-dimensional structure from sequence with per-residue confidence estimates. A predicted model is a hypothesis with an attached uncertainty. *See also* **pLDDT**, **Predicted aligned error**. **Amino acid** : The monomer unit of a protein; twenty are encoded by the standard genetic code. Their side-chain chemistry is what substitution matrices attempt to summarise. **Annotation** : The assignment of coordinates and biological meaning to features in a sequence. Divided into structural annotation, which says where features are, and functional annotation, which says what their products do. **Assembly** : The reconstruction of a genome sequence from overlapping reads, and also the resulting set of contigs or chromosomes. Assembly quality is described by contiguity, completeness and correctness, which can vary independently. *See also* **Contig**, **BUSCO**. **Asymmetric unit** : The smallest part of a crystal from which the whole can be generated by symmetry, and what a PDB coordinate file actually contains. It need not correspond to the biological assembly, so three protein chains in a file may be one crystallographic arrangement rather than a trimer. **AUGUSTUS** : A gene prediction program for eukaryotic genomes that models exon–intron structure with a generalised hidden Markov model, and can be guided by transcript or protein evidence. ## B {#sec-glossary-b} **BAM** : The binary, compressed form of SAM. When coordinate-sorted and indexed it supports random access to any genomic region, which is why alignment files are distributed in this form. *See also* **SAM**, **CRAM**, **Index**. **Batch effect** : Systematic technical variation associated with when or where samples were processed, which can exceed the biological effect of interest. Adjustable when batch and biological group are crossed; irrecoverable when they are confounded. *See also* **Confounding**. **Benjamini–Hochberg procedure** : A step-up method that controls the false discovery rate by comparing ranked p-values against a proportionally increasing threshold. The default correction in genomics, where thousands of hypotheses are tested at once. *See also* **False discovery rate**. **B-factor** : A column of a PDB coordinate file reporting positional uncertainty or thermal motion for each atom in an experimentally determined structure. In a predicted model the same column usually holds a confidence score instead, which must not be read as a temperature factor. *See also* **pLDDT**. **BGZF** : Blocked GNU Zip Format: a gzip-compatible compression in which data are stored in independent blocks of at most 64 kb, each with a recorded virtual offset. This is what allows a compressed VCF or BAM to be indexed and read at a single genomic position. *See also* **Tabix**, **Index**. **Bioconda** : A community-maintained distribution channel supplying bioinformatics software as conda packages, which is how most command-line tools in this book are installed and pinned. *See also* **Conda**. **Bioconductor** : A coordinated collection of R packages for the analysis of high-throughput biological data, built around shared data structures so that packages interoperate. *See also* **SummarizedExperiment**. **Biopython** : A Python library providing sequence, alignment, structure and database objects, including the `SeqIO` and `PDB` modules used throughout this book. **Biostrings** : The Bioconductor class and package for biological sequences in R, providing memory-efficient storage and vectorised operations such as base composition, pattern matching and reverse complementation. *See also* **Bioconductor**. **Bit score** : An alignment score normalised so that it is independent of the scoring system and of the database searched, which is why it is comparable between searches while an E-value is not. *See also* **E-value**. **BLAST** : The Basic Local Alignment Search Tool, a heuristic method that finds local alignments between a query and a sequence database by extending short exact matches, and reports each hit with an E-value. Fast enough to search an archive, and sensitive enough for most homology questions. *See also* **E-value**, **Local alignment**. **BLOSUM** : A family of amino acid substitution matrices estimated from aligned blocks of related proteins. The number names the identity level of the blocks used, so BLOSUM45 suits distant comparisons and BLOSUM80 close ones. *See also* **Scoring matrix**. **Broad consent** : A consent model in which participants permit future health-related research in general, with specific uses overseen by an ethics committee and a data access committee rather than approved individually. It keeps cohorts usable for decades at the cost of making the participant's agreement thin. *See also* **Tiered consent**, **Dynamic consent**. **BUSCO** : A method and a set of benchmarking gene collections used to estimate how complete an assembly or annotation is, by counting how many near-universal single-copy orthologues are present, duplicated, fragmented or missing. **BWA** : The Burrows–Wheeler Aligner, a short-read aligner that indexes the reference genome with a compressed suffix array; `bwa mem` is the standard mode for reads of 70 bp and longer. *See also* **Mapping quality**. ## C {#sec-glossary-c} **CDS** : Coding sequence: the portion of a transcript that is translated, beginning at a start codon and ending at a stop codon. In GFF3 the CDS feature is a child of a transcript feature. *See also* **Open reading frame**. **Central dogma** : The model of information flow in cells from DNA to RNA to protein, with replication maintaining the DNA copy. Each step explains why a particular representation exists in bioinformatics. **Centrality** : Any measure of how important a node is in a network. Degree centrality counts a node's edges and identifies hubs; betweenness centrality counts the shortest paths running through a node and identifies connectors between modules. *See also* **Hub**. **CIGAR** : The field of a SAM record describing how a read aligns to the reference as a run-length string of operations — match or mismatch, insertion, deletion, soft or hard clipping. The CIGAR, not the read length, determines where an alignment ends. *See also* **SAM**, **Soft clipping**. **ClinVar** : A public archive of submitted assertions about the clinical significance of human variants, with the evidence, the submitter and a review status recorded for each. The review status matters more than the significance label. *See also* **ACMG/AMP criteria**. **Clustering** : Grouping observations by similarity without using labels. The result depends on the distance measure, the linkage or algorithm and the number of groups requested, none of which the data determine on their own. *See also* **Unsupervised learning**. **Coding density** : The percentage of a genome covered by coding sequence. A useful annotation diagnostic: in compact bacterial genomes it approaches 95%, and a prediction far below the expected value has lost or fragmented genes. **Codon** : A triplet of nucleotides specifying one amino acid or a translation stop. **Codon usage bias** : The unequal use of synonymous codons within a genome, which differs between organisms and between highly and lowly expressed genes. Gene finders exploit it to distinguish coding from non-coding sequence. **Coercion** : Automatic conversion of values to a common type, which in R promotes a mixed vector to the most general type present, so one character value turns a whole numeric vector into text. A numeric column that will not do arithmetic has almost always been coerced. *See also* **Factor**. **Community detection** : Algorithms that partition a network into groups with more internal than external edges. They always return a partition, so the diagnostic question is whether the modularity is high and whether different algorithms agree. *See also* **Modularity**, **Module**. **Conda** : A package and environment manager that installs software and its dependencies into isolated environments described by a specification file, which is how this book pins its software. *See also* **Bioconda**, **Environment specification**. **Confounding** : The situation in which two possible explanations for a result cannot be separated because they vary together, as when every treated sample is sequenced in one run. No statistical adjustment repairs a confounded design. *See also* **Batch effect**. **Confusion matrix** : The table of predicted against true classes from which accuracy, precision, recall and specificity are all computed. Reporting it avoids the ambiguity of any single summary figure. *See also* **Precision**, **Sensitivity**. **Consent** : The permission under which data about a person may be collected and used. In genomics the practical question is not whether consent exists but whether it covers the analysis intended, and whether the restriction is enforced in the sample table and the pipeline. **Contact map** : A matrix in which entry (i, j) is set when residues i and j of a structure lie closer than a chosen cut-off, typically 8 Å between Cα atoms. Bands running perpendicular to the diagonal indicate antiparallel β-strands. **Contig** : A contiguous sequence assembled from overlapping reads. An assembly's contiguity is usually summarised by the number of contigs and by N50, the length such that half the assembly lies in contigs at least that long. **Controlled access** : A governance model in which individual-level data remain in a managed repository and investigators apply to a data access committee for permission to use them for a stated purpose. The standard arrangement for human genomic data, which cannot be anonymised. *See also* **dbGaP**, **European Genome-phenome Archive**, **Five safes**. **Counts per million** : A simple library-size normalisation in which each count is divided by the sample's total and multiplied by a million, abbreviated CPM. It corrects depth but not composition, which is why differential expression methods use more robust size factors. *See also* **Median-of-ratios normalisation**. **Coverage** : The average number of reads overlapping a position, also called depth. Coverage determines the confidence of a variant call, and uneven coverage is a diagnostic of capture bias, repeats or contamination. **CRAM** : A reference-based compressed alignment format that stores differences from the reference rather than full read sequences, giving smaller files than BAM at the cost of requiring the exact reference to decode. **Cross-validation** : Repeatedly fitting a model on part of the data and evaluating it on the held out remainder, to estimate performance on unseen data. Valid only if every fitting step, including feature selection and normalisation, happens inside the training part. *See also* **Overfitting**. **Curated database** : A resource whose records are reviewed, corrected and made non-redundant by expert or automated curation, such as UniProtKB/Swiss-Prot or RefSeq, as distinct from a primary archive that stores what was deposited. *See also* **Primary archive**. ## D {#sec-glossary-d} **Data access committee** : The body that assesses applications to use controlled-access data, judging whether the proposed use falls within participants' consent, whether the applicant's institution accepts the obligations, and whether the data requested are the minimum needed. **Data frame** : R's table type: a list of equal-length vectors presented as columns, each with its own type. *See also* **Tibble**. **Data management plan** : A document written before data exist that states what will be collected, where it will live, who may access it, how long it is kept, what is deposited at publication and who remains responsible afterwards. **Data sovereignty** : The principle that data about a people, or collected within a jurisdiction, is governed under that community's or country's authority rather than that of whoever holds a copy. *See also* **Helicopter research**. **dbGaP** : The database of Genotypes and Phenotypes at NCBI, which holds controlled access human study data with openly discoverable study-level metadata. **dbSNP** : An NCBI archive of short genetic variation, which assigns the `rs` identifiers used to refer to variants across resources. An `rs` number asserts that a variant was observed, not that it is common or benign. **Degree** : The number of edges attached to a node in a network. *See also* **Centrality**, **Hub**. **De-identification** : The removal of direct identifiers from a dataset. It guarantees only that those fields are gone; it does not make genomic data anonymous, because the genotype is itself an identifier. *See also* **Pseudonymisation**. **DESeq2** : An R/Bioconductor package for differential expression analysis of count data, which models counts with a negative binomial distribution and shrinks dispersion and fold-change estimates towards a fitted trend. *See also* **edgeR**, **Differential expression**. **Differential expression** : The comparison of transcript or gene abundance between conditions, testing for each gene whether the difference exceeds what the measurement's noise model allows. Requires normalisation, a replicated design and multiple-testing control. **Dimensionality reduction** : Projection of high-dimensional data onto a few axes for visualisation or downstream modelling, as in principal component analysis, t-SNE and UMAP. Separation in such a projection is a hypothesis, not a test result. **Dispersion** : In a negative binomial model of count data, the parameter describing how much more variable the counts are than a Poisson process would predict. With few replicates it is estimated poorly per gene, so DESeq2 and edgeR shrink per-gene estimates towards a fitted mean–dispersion trend. *See also* **DESeq2**. **DNA** : Deoxyribonucleic acid, the double-stranded polymer of nucleotides that stores hereditary information. **Domain** : A structurally or functionally autonomous unit of a protein, which often folds independently and recurs across proteins. Domain content is the usual basis for functional annotation by similarity. *See also* **Pfam**, **InterPro**. **dplyr** : The tidyverse package providing the data-manipulation verbs `filter`, `select`, `mutate`, `arrange`, `group_by` and `summarise`, composed with a pipe. *See also* **Pipe**, **Tidy data**. **DSSP** : The standard algorithm for assigning secondary structure to a set of coordinates from backbone hydrogen-bonding patterns. *See also* **Secondary structure**. **Dual-use research of concern** : Research conducted for legitimate purposes that could be directly misapplied to cause harm, such as work that increases a pathogen's transmissibility or that reconstructs a dangerous agent from published sequence. **Duplicate marking** : Flagging read pairs with identical alignment start positions as likely products of the same original fragment, so that variant callers do not count one molecule many times. **Dynamic consent** : A consent model in which participants hold an account through which they can see what has been done with their data and change their permissions at any time. The most responsive model and the most demanding to operate. ## E {#sec-glossary-e} **edgeR** : An R/Bioconductor package for differential expression of count data using negative binomial models and empirical Bayes moderation of dispersion estimates. *See also* **DESeq2**. **Enrichment analysis** : Testing whether a gene set is over-represented in a category — a Gene Ontology term, a pathway — relative to a defined background. The background must be the set of genes the experiment could have detected, not the whole genome. *See also* **Gene Ontology**, **Multiple testing**. **Ensembl** : A genome annotation resource and browser for vertebrate and other genomes, providing gene models, comparative genomics and variation data through a browser, an API and bulk files. **Environment specification** : A file listing the exact software packages, versions and channels an analysis requires, from which the environment can be recreated. A version range is not a specification. **European Genome-phenome Archive** : A controlled-access archive, operated by EMBL-EBI and the Centre for Genomic Regulation, for human data whose consent does not permit open distribution. Abbreviated EGA. *See also* **Controlled access**. **E-value** : In a database search, the number of alignments with at least the observed score expected to occur by chance given the database size. Because it accounts for length and search space, it is a better guide to homology than percentage identity. *See also* **BLAST**. **Exome** : The protein-coding portion of a genome, about 1–2% of human sequence, and by extension the capture-based sequencing strategy targeting it. *See also* **Whole-exome sequencing**. **Exon** : A segment of a transcript retained after splicing. Exon boundaries are why a transcript's coordinates differ from its gene's genomic coordinates. ## F {#sec-glossary-f} **Factor** : R's type for a categorical variable, stored as integer codes plus a table of permitted levels. Because the codes are integers, arithmetic on a factor silently uses them, and only `as.numeric(as.character(x))` recovers the labels. **FAIR principles** : A set of expectations that data be findable, accessible, interoperable and reusable, through persistent identifiers, documented access procedures, standard formats and vocabularies, and clear licensing. FAIR does not mean open: controlled-access data can be fully FAIR. **False discovery rate** : The expected proportion of false positives among the tests declared significant. Controlling it, rather than the probability of any false positive, is what makes genome-wide testing tractable. *See also* **Benjamini–Hochberg procedure**, **Multiple testing**. **FASTA** : A plain-text format in which each record is a header line beginning with `>` followed by sequence lines. It carries sequence and a free-text description and nothing else — no quality, no coordinates, no annotation. **FASTQ** : The read format that pairs each sequence with per-base quality scores, in four lines per record. The quality line is what distinguishes it from FASTA and what quality control operates on. *See also* **Phred quality score**. **FastQC** : A program that computes per-base and per-read quality summaries for sequencing data and presents them as a report. Its red and amber flags are heuristics tuned for whole-genome data and must be read with the library type in mind. **featureCounts** : A program that assigns aligned reads to genomic features such as exons or genes, producing the count matrix used in differential expression analysis. **Five safes** : A framework that decomposes safe data access into five separable questions — safe projects, safe people, safe settings, safe data and safe outputs — so that weakness in one dimension can be compensated in another. **FLAG** : The second field of a SAM record: a bit field encoding whether the read is paired, properly paired, mapped, reverse-stranded, the first or second mate, a duplicate, or a secondary or supplementary alignment. *See also* **SAM**, **Proper pair**. **flagstat** : The `samtools` subcommand that counts the alignment records in a BAM file by FLAG category — total, mapped, properly paired, singletons, duplicates — and the usual first summary of an alignment. *See also* **SAMtools**, **FLAG**. **FM-index** : A compressed full-text index built on the Burrows–Wheeler transform that supports exact substring search in space close to that of the compressed text. It is what allows a short-read aligner to index a whole genome and place millions of reads against it. *See also* **BWA**. **Frameshift** : An insertion or deletion whose length is not a multiple of three, which changes the reading frame of every downstream codon and usually introduces a premature stop. Generally more damaging than a single missense change. **Functional annotation** : Assignment of biological meaning — product name, domains, Gene Ontology terms, pathway membership — to a predicted feature, almost always by transfer from homologous sequences. *See also* **Structural annotation**, **Homology**. ## G {#sec-glossary-g} **GATK** : The Genome Analysis Toolkit, a suite for variant discovery from high-throughput sequencing data, including the local reassembly-based HaplotypeCaller and the filtering steps that follow it. **GC content** : The percentage of bases in a sequence that are G or C. It varies systematically between genomes and along them, affects sequencing coverage and gene-prediction models, and is the quickest compositional check for contamination. **GenBank** : The NCBI primary nucleotide sequence archive, and the flat-file format in which its records are distributed, carrying sequence together with a feature table. Part of the International Nucleotide Sequence Database Collaboration with ENA and DDBJ. *See also* **Primary archive**. **Gene** : A segment of DNA transcribed as a unit, together with the regulatory sequence governing it. The operational definition varies between resources, which is why gene counts differ between annotations of the same genome. **Gene Ontology** : A structured, species-independent vocabulary describing gene product molecular functions, biological processes and cellular components, with terms arranged in a hierarchy and annotations carrying evidence codes. **Gene set enrichment analysis** : A method, abbreviated GSEA, that tests whether members of a gene set are concentrated at one end of a ranked list of all genes, rather than testing a thresholded subset. It detects coordinated small shifts that over-representation analysis misses, at the cost of depending on the ranking statistic. *See also* **Enrichment analysis**. **Genetic code** : The mapping from codons to amino acids and stop signals. Variants of the standard code exist: NCBI translation table 4, used by *Mycoplasma* and relatives, reads TGA as tryptophan rather than as a stop, which breaks gene prediction run under the bacterial default. *See also* **Translation table**. **Genome browser** : An interactive viewer that displays annotation, alignment and variation tracks against genomic coordinates, such as Ensembl, the UCSC browser or IGV. Looking at the reads is the fastest way to reject a spurious variant call. **Genome-wide association study** : A study that tests association between many common variants and a phenotype across a population, correcting for multiple testing and population structure. Abbreviated GWAS. **GFF3** : A nine-column tab-separated annotation format giving each feature's sequence, type, coordinates, strand, phase and attributes, with `ID` and `Parent` attributes expressing the gene–transcript–CDS hierarchy. Coordinates are one-based and inclusive. *See also* **GTF**. **ggplot2** : The R implementation of the grammar of graphics, in which a plot is built from data, aesthetic mappings, geometries, scales, facets and a theme, added as layers. It requires data in long form. *See also* **Tidy data**. **Git** : A distributed version control system that records the history of a set of files as a series of commits. In an analysis project it holds code, configuration and documentation, never derived results or controlled-access data. *See also* **Version control**. **Global alignment** : An alignment of two sequences along their entire length, appropriate when they are expected to correspond end to end. Computed optimally by the Needleman–Wunsch algorithm. **gnomAD** : A reference database of allele frequencies aggregated from large exome and genome collections, used as the population-frequency evidence in variant interpretation. Its ancestry composition determines how much power that evidence has for a given patient. **GRanges** : The Bioconductor class for genomic intervals, each carrying a sequence name, start, end and strand plus arbitrary metadata columns, with set operations and overlap queries defined on it. *See also* **Bioconductor**. **GTF** : A tab-separated annotation format closely related to GFF, with a fixed attribute convention centred on `gene_id` and `transcript_id`. Widely expected by RNA sequencing tools. *See also* **GFF3**. ## H {#sec-glossary-h} **Haplotype** : A combination of alleles at linked sites inherited together on the same chromosome. *See also* **Linkage disequilibrium**. **Hardy–Weinberg equilibrium** : The expected relationship between allele and genotype frequencies in a large randomly mating population, which gives genotype frequencies $p^2$, $2pq$ and $q^2$. Departures are used as a quality filter on genotype data and as evidence of population structure. **Helicopter research** : Research in which investigators from a well-resourced institution collect samples or data in a lower-income setting, analyse them elsewhere, and leave no capacity, authorship or benefit behind. Also called parachute or extractive research. *See also* **Data sovereignty**. **Hidden Markov model** : A probabilistic model of a sequence of observations generated by unobserved states with defined transition probabilities. Used for gene structure prediction, protein family profiles and secondary structure assignment. *See also* **HMMER**, **AUGUSTUS**. **HIPAA** : The US Health Insurance Portability and Accountability Act, which regulates protected health information held by covered entities such as providers and health plans and by their business associates. Many research laboratories are not covered entities, and genomic sequence is not among the identifier types its Safe Harbor standard enumerates. **HMMER** : A program suite that searches sequences against profile hidden Markov models and builds such profiles from alignments, giving greater sensitivity to distant homology than pairwise search. *See also* **Pfam**, **Hidden Markov model**. **Homology** : Shared ancestry between sequences or structures. Homology is inferred from similarity, never measured directly, and sequences are either homologous or not — "percentage homology" is a category error. *See also* **Orthologue**, **Paralogue**. **Homology modelling** : Prediction of a protein structure by aligning its sequence to that of an experimentally determined relative and copying the template's coordinates. Its accuracy is governed mainly by sequence identity to the template, and it cannot predict a fold absent from the database. **Hub** : A network node with unusually high degree. Hubs are often essential, but in literature-derived networks they are also the best-studied proteins, which is a confound rather than a finding. *See also* **Degree**. **Hypergeometric test** : The exact test for over-representation of a category in a sample drawn without replacement, and the statistic underneath most enrichment tools: it asks how improbable the observed overlap between a gene list and a term would be if the list were drawn at random from the background universe. *See also* **Enrichment analysis**. ## I {#sec-glossary-i} **Incidental finding** : A result of potential clinical significance discovered outside the question a study set out to answer. Whether to return one should be decided in the protocol, before sequencing, and by the people with clinical and ethical standing to decide. *See also* **Secondary finding**. **Indel** : An insertion or deletion of one or more bases, reported as a single variant class because alignment often cannot distinguish which event occurred. *See also* **Frameshift**. **Index** : An auxiliary file that maps coordinates or identifiers to byte offsets in a large data file so that a region can be retrieved without reading the whole file. Required for BAM, CRAM, compressed VCF and indexed FASTA. *See also* **Tabix**. **Informed consent** : The process by which a participant, having been given an adequate explanation of a study's purposes, procedures, risks and data uses, voluntarily agrees to take part. *See also* **Broad consent**. **Insert size** : The length of the original DNA fragment from which a read pair was sequenced, reported in SAM as the signed template length. Its distribution is a property of library preparation, and an implausibly tight one is a sign of simulated data. **InterPro** : An integrative resource that combines protein family, domain and site signatures from several member databases, and the `InterProScan` tool that searches a sequence against all of them at once. *See also* **Pfam**, **Domain**. **Intron** : A segment of a primary transcript removed by splicing. Introns are why eukaryotic gene prediction is substantially harder than bacterial gene prediction. **Isoform** : One of several distinct transcripts or protein products of the same gene, typically arising from alternative splicing or alternative promoters. A protein record's canonical isoform may differ from the transcript you translated. ## K {#sec-glossary-k} **KEGG** : A collection of manually drawn pathway maps linking genes and metabolites to molecular interaction and reaction networks, widely used as the pathway vocabulary for enrichment analysis. **k-means clustering** : An algorithm that partitions observations into a pre-specified number of clusters by iteratively assigning each to the nearest centroid and recomputing centroids. The number of clusters is an input, not a result. *See also* **Clustering**. **k-mer** : A substring of length *k*. Counting k-mers underlies assembly graph construction, sequence comparison without alignment, and contamination screening. ## L {#sec-glossary-l} **Leakage** : Any route by which information from held-out data influences model fitting — most often feature selection, normalisation or hyperparameter tuning performed before the data were split. It inflates reported performance, and label permutation is the standard test for it. *See also* **Cross-validation**. **Licence** : The terms under which others may use a work. Code, text and data follow different conventions — permissive and copyleft software licences, Creative Commons for text and figures, open data licences for databases — and unlicensed code is by default not reusable. **Linkage disequilibrium** : The non-random association of alleles at nearby loci, arising because recombination has not yet separated them. It makes genome-wide association possible, by letting a typed marker tag an untyped causal variant, and it is also why polygenic scores transfer poorly between populations with different haplotype structure. *See also* **Polygenic score**. **Local alignment** : An alignment of the best-scoring subsequences of two sequences, appropriate when only part of each is expected to correspond. Computed optimally by the Smith–Waterman algorithm and heuristically by BLAST. **Long-read sequencing** : Sequencing technologies producing reads of thousands to millions of bases, at a higher per-base error rate than short reads, which resolve repeats and structural variation that short reads cannot. ## M {#sec-glossary-m} **MAFFT** : A program for multiple sequence alignment offering a range of algorithms from fast progressive methods for large inputs to accurate iterative refinement for small ones. **Makefile** : A file declaring, for each output, the inputs and the command that produces it, from which `make` computes what is out of date and rebuilds only that. It records dependencies rather than execution order. *See also* **Workflow manager**. **MA plot** : A scatter plot of log fold change against mean expression used to inspect a differential expression result. A systematic vertical shift indicates a normalisation problem rather than biology. **Mapping quality** : A read's alignment confidence, expressed on the Phred scale, which reports the probability that the read came from somewhere else in the reference. Low values indicate repeats or paralogues and are a standard variant filter. **MAPQ** : The mapping quality field of a SAM record, a Phred-scaled estimate of the probability that the read was placed in the wrong location. A MAPQ of 0 means the read aligns equally well elsewhere, usually in a repeat, and says nothing about read quality. *See also* **Mapping quality**, **SAM**. **Median-of-ratios normalisation** : The size-factor estimator used by DESeq2: each sample's factor is the median across genes of its count divided by the geometric mean of that gene across samples. It is robust to a minority of strongly changing genes in a way that total-count scaling is not. *See also* **Counts per million**, **DESeq2**. **Minor allele frequency** : The frequency of the less common allele at a site, abbreviated MAF. Rare variants carry little individual-identifying information whereas common ones carry a great deal, which is why forensic and sample-tracking panels use common markers. *See also* **Allele frequency**. **Missense variant** : A single-nucleotide change that alters the encoded amino acid. Its consequence depends on the residue's structural and functional context, which is why prediction of missense effect remains imprecise. **mmCIF** : The macromolecular Crystallographic Information File format, the current standard representation for structures in the Protein Data Bank, which unlike the legacy PDB format is not limited by fixed column widths. *See also* **PDB**. **Modularity** : A score, written *Q*, measuring how many more edges fall inside a proposed partition's groups than would be expected at random. Low modularity across several algorithms indicates that the communities returned are artefacts of the method. *See also* **Community detection**. **Module** : A densely interconnected group of nodes in a network, often corresponding to a protein complex or a pathway. *See also* **Clustering**. **Multiple sequence alignment** : An alignment of three or more sequences, abbreviated MSA, from which conserved columns, phylogenies and profile models are derived. Progressive methods build it from pairwise alignments, so an early error propagates. *See also* **MAFFT**. **Multiple testing** : The inflation of false positives that occurs when many hypotheses are tested at once, and the corrections that control it. A per-test p-value threshold of 0.05 across 20,000 genes expects 1,000 false positives. *See also* **False discovery rate**. **Mutation** : A change in a DNA sequence relative to a reference or a parental sequence. In clinical contexts "variant" is now preferred, because "mutation" carries an implication of pathogenicity that the observation does not support. **Mutational signature** : The characteristic distribution over mutation types, conventionally 96 classes of substitution in trinucleotide context, that a particular mutational process generates. A tumour's spectrum is modelled as a mixture of signatures; the result describes chemistry, not functional consequence. ## N {#sec-glossary-n} **N50** : The length such that half of an assembly lies in contigs at least that long. It measures contiguity rather than correctness and rewards over-aggressive assembly, so it should be reported with total length, contig count and a completeness estimate. *See also* **Contig**, **BUSCO**. **Nagoya Protocol** : An international agreement under the Convention on Biological Diversity governing access to genetic resources and the fair and equitable sharing of benefits from their use, implemented through prior informed consent from the provider country and mutually agreed terms. Its application to digital sequence information remains under negotiation. **NCBI** : The US National Center for Biotechnology Information, which operates GenBank, RefSeq, PubMed, dbSNP, ClinVar, the Sequence Read Archive and the BLAST services. **NDPR** : The Nigeria Data Protection Regulation of 2019, which established consent requirements, data subject rights and cross-border transfer restrictions for personal data in Nigeria; the Nigeria Data Protection Act 2023 placed the regime on a statutory footing with a dedicated commission as regulator. **Needleman–Wunsch algorithm** : The dynamic programming algorithm that computes an optimal global alignment of two sequences under a given scoring scheme. *See also* **Global alignment**. **Next-generation sequencing** : The massively parallel sequencing technologies that replaced capillary sequencing, producing millions of short reads per run. Abbreviated NGS. **Nonsense-mediated decay** : A surveillance pathway that degrades transcripts bearing a premature termination codon, which is why many nonsense and frameshift variants reduce protein to near zero rather than producing a truncated product. **Nonsense variant** : A single-nucleotide change that converts an amino acid codon into a stop codon, truncating the protein. *See also* **Nonsense-mediated decay**. **Normalisation** : Adjustment of measurements so that they are comparable across samples, for instance correcting sequencing counts for differences in library size and composition. A raw count is not an abundance. **Nucleotide** : The monomer unit of DNA and RNA, comprising a base, a sugar and a phosphate. ## O {#sec-glossary-o} **Open reading frame** : A stretch of sequence in one reading frame that contains no stop codon, conventionally from a start codon to the next in-frame stop. An ORF is a candidate gene, not a gene. *See also* **Ab initio gene prediction**. **Orthologue** : A homologous sequence in another species that descends from the same ancestral sequence by speciation, and which usually retains the ancestral function. *See also* **Paralogue**. **Overfitting** : Fitting the noise in the training data rather than the signal, so that performance on new data is much worse than on the data used to build the model. The main risk when features greatly outnumber samples. *See also* **Cross-validation**. ## P {#sec-glossary-p} **Paired-end sequencing** : Sequencing both ends of each fragment, giving two reads of known relative orientation and approximate separation, which improves alignment in repetitive regions and allows detection of structural variation. **Paralogue** : A homologous sequence within the same genome that arose by duplication. Paralogues may have diverged in function, so transferring annotation between them is less safe than between orthologues. **Pathogenicity** : The degree to which a variant causes disease. Classified from weighed evidence rather than measured, and revised as evidence accumulates. *See also* **ACMG/AMP criteria**. **PDB** : The Protein Data Bank, the archive of experimentally determined macromolecular structures, and the legacy fixed-column coordinate file format it introduced. *See also* **mmCIF**, **B-factor**. **Permutation test** : A test in which the labels or the data are shuffled many times to build the null distribution of a statistic empirically. In machine learning it is the standard check for leakage: with labels permuted, an honest pipeline must perform no better than chance. *See also* **Leakage**. **Pfam** : A database of protein families represented as profile hidden Markov models built from curated seed alignments, now distributed as part of InterPro. *See also* **HMMER**. **Pharmacogenomics** : The study of how inherited variation affects drug response, and the clinical application of that knowledge through gene–drug guidelines covering dosing and drug choice. **Phred quality score** : A base-calling confidence expressed as $-10\log_{10}$ of the error probability, so that 20 means one error in 100 and 30 means one in 1,000. Encoded as ASCII characters in FASTQ. *See also* **FASTQ**. **Phylogenetics** : The inference of evolutionary relationships from molecular data, usually presented as a tree whose topology, branch lengths and support values each carry different information. **Pileup** : A per-position view of the bases and qualities of all reads covering a site, and the input to most variant callers. Inspecting the pileup is the fastest way to reject an artefactual call. **Pipe** : An operator that passes the output of one expression as the input of the next: `|` in the Unix shell, `|>` in base R since version 4.1 and `%>%` in the magrittr and tidyverse idiom. It makes a sequence of transformations readable in the order they happen. **Pipeline** : An ordered set of analysis steps, each consuming the previous step's output. Written as a workflow definition rather than as a list of commands, it becomes reproducible and restartable. **pLDDT** : The per-residue confidence score of an AlphaFold model, on a 0–100 scale, stored in the B-factor column. Low values often indicate genuine disorder rather than prediction failure. *See also* **AlphaFold**. **Polygenic score** : A weighted sum of many variant dosages used to estimate an individual's genetic liability to a trait. Weights estimated in one population transfer poorly to others because the markers tag causal variants through population-specific linkage disequilibrium. **Precision** : The proportion of positive predictions that are correct. Reported alongside sensitivity, because either alone can be made perfect by a trivial predictor. *See also* **Sensitivity**. **Precision–recall curve** : A plot of precision against recall across classification thresholds, and the appropriate summary when the positive class is rare, because unlike a receiver operating characteristic curve it is not flattered by a large number of true negatives. *See also* **ROC curve**, **Precision**. **Predicted aligned error** : A matrix accompanying a predicted structure giving the expected positional error of residue *i* when the model is superposed on residue *j*. It is the quantity to consult before trusting the relative placement of two domains. *See also* **AlphaFold**. **Primary archive** : A repository that stores records as deposited, with the submitter responsible for their content and earlier versions retained, such as GenBank or the Sequence Read Archive. *See also* **Curated database**. **Principal component analysis** : A linear projection onto orthogonal axes ordered by the variance they capture, used to visualise structure and to summarise correlated features. The components are directions of variance, not of biological interest. *See also* **Dimensionality reduction**. **Prodigal** : An *ab initio* gene prediction program for prokaryotic genomes, which trains a coding model on the genome itself and scores open reading frames by codon composition and ribosome binding site signals. Its default translation table is the bacterial standard, which is wrong for some clades. *See also* **Translation table**. **Prokka** : A pipeline that wraps gene prediction with RNA gene finding and functional annotation by homology to produce a complete prokaryotic annotation in standard formats. **Proper pair** : A read pair that the aligner considers correctly placed: both mates mapped, in the expected relative orientation, and separated by a plausible insert size. The proportion of properly paired reads is a standard alignment diagnostic. *See also* **flagstat**. **Pseudogene** : A sequence recognisably derived from a gene but no longer producing a functional product, through truncation, frameshift or loss of expression. Pseudogene loci are where annotation methods most often disagree. **Pseudonymisation** : Replacement of direct identifiers with a code while retaining the ability to re-identify through a separately held key. Pseudonymised data remain personal data in law. *See also* **De-identification**. **p-value** : The probability of observing data at least as extreme as those obtained if the null hypothesis were true. It is not the probability that the hypothesis is false, and it says nothing about effect size. ## Q {#sec-glossary-q} **Quality control** : The checks applied to data before analysis to establish that they are fit for the question, and the diagnostics that detect contamination, bias or technical failure. The step whose omission costs the most time later. **QUAST** : A tool that computes assembly quality metrics — contiguity, length distribution, misassemblies against a reference — and reports them in a single comparison table. ## R {#sec-glossary-r} **R** : A language and environment for statistical computing and graphics, the usual choice in this book for modelling and figures, extended for biology by Bioconductor. **Radius of gyration** : The root-mean-square distance of a structure's atoms from its centre of mass, a single-number measure of compactness. For a model whose domains are joined by disordered linkers it describes one arbitrary arrangement rather than a property of the protein. **Ramachandran plot** : A scatter plot of the backbone dihedral angles φ and ψ for every residue in a structure. Residues cluster in the sterically allowed α-helical and β regions, and a point in a forbidden area is a modelling error until shown otherwise. **Random forest** : An ensemble classifier or regressor built from many decision trees trained on bootstrap samples with random feature subsets, whose predictions are aggregated. Robust on tabular data and resistant, though not immune, to overfitting. **Read** : A single sequence produced by a sequencing instrument, with per-base quality scores. A read is an observation; where it came from is an inference. **Recycling** : R's rule of repeating the shorter operand of a vectorised operation until the lengths match, without warning when they divide evenly. On a matrix it recycles down the columns, so dividing a gene-by-sample count matrix by a vector of library sizes scales the wrong dimension and returns a plausible, wrong answer. **Reference genome** : An assembled, coordinate-stable sequence used as the frame against which reads are aligned and variants reported. Results are only comparable between analyses using the same build, which is why the build identifier belongs in every methods section. **RefSeq** : The NCBI curated, non-redundant collection of genomic, transcript and protein sequences, whose identifiers carry prefixes such as `NM_` for mRNA and `NP_` for protein. *See also* **Curated database**. **Reproducibility** : The property that another person, given the same data, can obtain the same numbers. In practice it requires code under version control, a pinned environment, recorded parameters and seeds, and documented data provenance. **Resolution** : For a crystal structure, the smallest spacing of diffraction data used, in ångströms, which bounds the detail the model can support: individual atoms near 1.5 Å, backbone and large side chains only beyond 3.5 Å. *See also* **R factor**. **R factor** : The disagreement between the diffraction amplitudes observed and those calculated from a crystallographic model. R-free is the same quantity computed on reflections withheld from refinement, and is the meaningful one because it cannot be improved by overfitting. **Ribosome binding site** : A sequence motif upstream of a start codon, in bacteria the Shine–Dalgarno sequence, which positions the ribosome. Gene finders model it to choose between candidate start codons. **RMSD** : Root-mean-square deviation: the square root of the mean squared distance between corresponding atoms of two superposed structures. Meaningless without the number of atoms and the selection used. *See also* **Superposition**. **RNA** : Ribonucleic acid, the single-stranded nucleotide polymer that carries genetic information from DNA to the ribosome and performs catalytic and regulatory roles. **RNA sequencing** : Measurement of transcript abundance by sequencing cDNA derived from RNA, abbreviated RNA-seq, quantified either by counting reads over features or by assigning them to transcripts probabilistically. *See also* **Differential expression**. **ROC curve** : A plot of true positive rate against false positive rate across classification thresholds, summarised by the area underneath it. On a strongly imbalanced problem it overstates performance. *See also* **Precision–recall curve**. ## S {#sec-glossary-s} **SAM** : The Sequence Alignment/Map format, a tab-separated text representation of read alignments carrying position, mapping quality, a CIGAR string describing the alignment, mate information and optional tags. *See also* **BAM**, **CRAM**. **SAMtools** : A program suite for manipulating alignment files — converting, sorting, indexing, filtering and computing pileups — and the companion to BCFtools for variant data. **Scoring matrix** : A table of log-odds scores for aligning each pair of residues, encoding how likely that substitution is between homologues relative to chance. *See also* **BLOSUM**. **Secondary finding** : A result of clinical significance that a project deliberately looks for beyond its primary question, as distinct from an incidental finding encountered unsought. *See also* **Incidental finding**. **Secondary structure** : Local, regularly hydrogen-bonded elements of a polypeptide chain, chiefly α-helices and β-strands, assigned from coordinates by algorithms such as DSSP. **Sensitivity** : The proportion of true positives that are detected, also called recall. In the annotation comparison of this book it is the proportion of reference coding sequences a prediction recovers. *See also* **Precision**. **Sequence Read Archive** : The NCBI primary archive of raw sequencing reads, abbreviated SRA, mirrored by the European Nucleotide Archive. **Shine–Dalgarno sequence** : A purine-rich ribosome binding motif a few bases upstream of a bacterial start codon, complementary to the 3' end of the 16S ribosomal RNA. Gene finders model it to choose between candidate start codons. *See also* **Ribosome binding site**. **Short-read sequencing** : Sequencing producing reads of roughly 50–300 bases at low per-base error rate, the basis of most resequencing, exome and RNA sequencing work. **Smith–Waterman algorithm** : The dynamic programming algorithm that computes an optimal local alignment of two sequences. *See also* **Local alignment**. **Snakemake** : A Python-based workflow manager that defines rules with input and output patterns, resolves the dependency graph, and executes it locally or on a cluster with environment isolation. *See also* **Makefile**, **Workflow manager**. **SNP** : A single-nucleotide polymorphism: a single-base difference between chromosomes that is common in a population. A single-base difference without the frequency claim is better called a single-nucleotide variant. **Soft clipping** : A CIGAR operation marking bases at the end of a read that are present in the record but not aligned, usually adapter sequence or accumulated errors. Hard-clipped bases, by contrast, are removed from the record entirely. *See also* **CIGAR**. **Specificity** : The proportion of true negatives correctly identified. Required alongside sensitivity whenever classes are unbalanced. **Splicing** : Removal of introns from a primary transcript and joining of exons, which is why transcript and genomic coordinates differ and why alignment of RNA reads to a genome must be splice-aware. **Standard streams** : The three channels every Unix program has: standard input, standard output and standard error. Keeping results on standard output and messages on standard error is what allows programs to be joined by pipes without corrupting the data. *See also* **Pipe**. **STRING** : A database of known and predicted protein–protein functional associations, integrating evidence channels into a confidence score. An edge asserts functional association, which is weaker than a physical interaction. **Structural annotation** : Determination of where features lie in an assembly — genes, exons, RNA genes, repeats — as distinct from what their products do. *See also* **Functional annotation**. **SummarizedExperiment** : The Bioconductor container holding one or more assay matrices together with sample and feature annotation, guaranteeing that subsetting cannot desynchronise them. *See also* **Bioconductor**. **Superposition** : Rigid-body rotation and translation of one structure onto another to minimise the distance between corresponding atoms. The choice of atoms superposed determines what the resulting deviations mean. *See also* **RMSD**. **Supervised learning** : Fitting a model to predict a known label or value from features, and evaluating it on data withheld from every stage of fitting. *See also* **Unsupervised learning**, **Cross-validation**. **Synonymous substitution** : A nucleotide change that leaves the encoded amino acid unaltered, possible because the genetic code is degenerate. Synonymous changes are not necessarily neutral: they can affect splicing, stability and translation rate. ## T {#sec-glossary-t} **Tabix** : An indexing method for position-sorted, block-compressed tab-delimited files such as VCF and GFF, which enables retrieval of a genomic interval without reading the file. *See also* **Index**. **Tibble** : The tidyverse's data frame, differing in three safety-relevant behaviours: no partial matching of column names, no dropping to a vector on single-column selection, and no silent conversion of character columns to factors. *See also* **Data frame**. **Tidy data** : A table in which each variable is a column, each observation a row and each cell a value. It is the shape `ggplot2` and grouped summaries require, and the opposite of the wide matrix that Bioconductor assays use; `pivot_longer` and `pivot_wider` convert between them. **Tiered consent** : A consent model in which participants decide separately about distinct uses — commercial research, international deposition, re-contact, return of results. It respects differing preferences and fragments the cohort into subsets whose permissions must be enforced in code. **Transcription** : Synthesis of RNA from a DNA template by RNA polymerase. **Transcriptome** : The complete set of transcripts in a cell or sample, and their abundances, at a given time. **Transfer RNA** : The adaptor RNA that carries an amino acid and recognises a codon through its anticodon, abbreviated tRNA. Predicted in genome annotation by specialised tools rather than by protein gene finders. **Translation** : Synthesis of a polypeptide from a messenger RNA template by the ribosome, reading the sequence in triplets. **Translation table** : One of the numbered NCBI genetic codes specifying which codons encode which amino acids and which terminate translation. Table 11 is the bacterial standard; table 4 applies to *Mycoplasma* and relatives; the table must be set before gene prediction. *See also* **Genetic code**. **Trimming** : Removal of adapter sequence and low-quality bases from reads before alignment. Over-trimming discards real signal, so trimming parameters belong in the methods section. *See also* **Adapter**. **t-SNE** : t-distributed stochastic neighbour embedding, a non-linear projection that preserves local neighbourhood structure for visualisation. Distances between clusters in a t-SNE plot carry no reliable meaning. *See also* **UMAP**, **Dimensionality reduction**. **Twilight zone** : The band of roughly 20–35% sequence identity over a protein-length alignment in which real homologues and chance similarities both occur, so identity alone cannot separate them and a significance estimate is required. *See also* **E-value**. ## U {#sec-glossary-u} **UMAP** : Uniform manifold approximation and projection, a non-linear dimensionality reduction method that preserves more global structure than t-SNE and runs faster, but whose output still depends strongly on its parameters. **UniProtKB** : The central protein sequence and annotation resource, divided into the manually reviewed Swiss-Prot section and the automatically annotated TrEMBL section, with every annotation carrying an evidence attribution. **Unix** : The family of operating systems, including Linux and macOS, whose shell and small composable text-processing programs form the working environment of bioinformatics. **Unsupervised learning** : Finding structure in data without labels, as in clustering and dimensionality reduction. Its output is a hypothesis about structure, not a test of one. *See also* **Supervised learning**. **Untranslated region** : The portion of a mature transcript upstream or downstream of the coding sequence, abbreviated UTR, which carries regulatory elements affecting stability and translation. ## V {#sec-glossary-v} **Variance-stabilising transformation** : A transformation of count data that makes variance approximately independent of the mean, so that distance-based methods such as clustering and principal component analysis are not dominated by the most highly expressed genes. **Variant** : A difference between an observed sequence and a reference at a defined position. The term is deliberately neutral about consequence. *See also* **Mutation**. **Variant calling** : Inference of genotypes at positions where reads differ from the reference, followed by filtering to remove systematic artefacts. A call is an inference about a genotype, with a confidence, not a measurement. **Variant of uncertain significance** : A variant whose available evidence is insufficient to classify as pathogenic or benign, abbreviated VUS. The classification describes the state of knowledge, which is why reclassification is routine and why uncertain results are commoner for under-represented populations. *See also* **ACMG/AMP criteria**. **VCF** : The Variant Call Format, a tab-separated text format with a header of metadata definitions and one line per variant site, carrying position, alleles, quality, filter status, an extensible `INFO` field and per-sample genotypes. Compressed and indexed for random access. *See also* **Tabix**. **Vectorisation** : Applying an operation to a whole array at once rather than element by element in a loop, which is both faster and less error-prone because no index is written by hand. It is the normal style in both NumPy and R. **Version control** : Systematic recording of changes to files over time so that any previous state can be recovered and every change attributed. *See also* **Git**. **Volcano plot** : A scatter plot of statistical significance against effect size for every gene in a differential expression analysis, used to show that the genes reaching significance also change by a worthwhile amount. ## W {#sec-glossary-w} **Whole-exome sequencing** : Sequencing of the captured protein-coding portion of a genome, abbreviated WES, which concentrates coverage where interpretable variants are most dense at the cost of missing regulatory and structural variation. **Whole-genome sequencing** : Sequencing of the entire genome without target capture, abbreviated WGS, giving more uniform coverage and access to non-coding and structural variation at greater cost per sample. **Workflow manager** : Software that executes an analysis from a declarative description of its steps and their dependencies, rerunning only what is out of date and recording what was run. *See also* **Makefile**, **Snakemake**. ## Y {#sec-glossary-y} **Y-STR** : A short tandem repeat marker on the Y chromosome. Because Y haplotypes and patrilineal surnames are inherited together, Y-STR profiles recovered from sequence data can be matched against genealogy databases to infer a surname. ## Z {#sec-glossary-z} **Z-score** : A value expressed as the number of standard deviations from the mean of its distribution. Used to standardise features before clustering and to put measurements on comparable scales, and meaningful only if the reference distribution is stated.