Bioinformatics and Computational Biology: Key Terms and Glossary

Prediction of genes from sequence signals and statistics alone, without protein or transcript evidence. The predictors disagree, and an unsupported call is a hypothesis, not a measurement.

The stable name of a database record. Alone, it does not name a sequence, because the sequence can change under a new version.

A gap cost with one charge for opening and a smaller charge for each further residue, so that one long indel is cheaper than many separate single-residue gaps.

The share of reads at a site that support an allele. In diploid germline DNA a real heterozygote is near one half, and one disagreeing read is still compatible with error.

Similarity without common ancestry. Failing to support homology is a different result, and sequences rarely prove analogy.

The same alignment table in compressed form. These notes do not specify the binary; it is converted when a person must look.

How common the true class is. A test with impressive sensitivity and specificity still yields mostly false positives when the class is rare, which is the usual situation for variants and for differential genes.

A technical group, such as the day of preparation, that can be perfectly aligned with the biological treatment. No false-discovery rule removes a contrast that is really the batch.

A procedure that sorts p values and keeps them up to the largest rank whose p value is at most that rank's share of a chosen false-discovery rate. Passing the cutoff is not 95% certainty on each gene.

In these notes, the job of making biological sense of sequence and structure data: storing them, comparing them, annotating them, and deciding what a similarity is allowed to say.

An alignment score from which the matrix and gap constants have been removed, so it can be compared when the database changes. Doubling the database doubles the E-value and leaves the bit score unchanged.

A family of matrices counted directly in conserved ungapped blocks after very similar sequences have been clustered. A higher BLOSUM number is a shorter evolutionary distance, the opposite direction from PAM.

The BLOSUM matrix whose clustering threshold was 62%. It does not mean 'use this when the sequences are 62% identical', and it is a matrix, not an identity.

A resampling of alignment columns with replacement, counting how often a clade reappears. It asks whether many columns agree, and a bootstrap percentage is not the probability that the clade is historically true.

A fossil or a dated biogeographic event, with its own uncertainty, used to place years on a clock. An error in where that calibration sits can leave a narrow interval around a date that is still wrong.

The string of alignment operations on a placed read. An M consumes one read base and one reference base without promising that they match, and an N skips reference as an intron rather than as a deletion.

Treating a label as confirmed because it is found in a database that received the label from the same kind of transfer. The circle is not an experiment.

One position in a multiple alignment, claiming that the letters in it are homologous. A tree, a logo, and a homology model all inherit that claim, and no later statistic repairs letters that do not belong together.

The wider use of quantitative models of living systems, including population genetics, physical simulation, and pathway kinetics. Sequence work is one part of that field, not a rival to it.

A preference for a column that a library of pairwise alignments agrees on, so support can pass from A-with-B and B-with-C to A-with-C.

Sequence from the wrong organism, often human or vector, inside an assembly. The alignment can be perfect and the biology still false; too much identity to the wrong species is the signature.

A stretch the assembler will write without a gap. It is unbroken sequence, not a claim that every join inside it is biologically correct.

Redundancy, the total bases sequenced divided by the genome length. It is not the fraction of the genome seen, and a high value does not bridge a repeat longer than the fragment.

The release or download date of the database that was searched. An E-value cannot be compared next year without it, because the database will have grown.

The DNA Data Bank of Japan, the third partner in the same exchanged nucleotide collection.

An assembly graph in which a k-mer is a node and an overlap of k−1 letters is an edge, used because short reads are too numerous to compare all against all. Repeats make the path non-unique.

The silent profile-HMM state that skips a consensus column, corresponding to a deletion relative to the family.

A single number summarising the difference between two sequences. Building a tree from distances discards which sites changed.

A search over orientations of two structures, scored by shape, charge, and statistical preferences. The score is not a dissociation constant, not proof the proteins meet in the cell, and not a drug lead.

Filling a score matrix so that each cell is decided from a few smaller cells, guaranteeing the optimal alignment under the chosen scores at a cost that grows with the product of the lengths.

How many alignments of at least the observed score a chance model expects for this query in a database of this size. It is not the bit score, it grows when the database grows, and it is not the probability that the sequences are homologous.

The European Nucleotide Archive, redundant with GenBank and DDBJ on purpose, and equally not a reviewer of function or species.

The label on a Gene Ontology annotation that says whether the support was an assay, a mutant, a similarity, an electronic transfer, or a literature statement. Most annotations in a large database are electronic copies.

Gene models supported by protein alignments or by transcript data. Where they conflict with an ab initio call, the evidence should win, and a gene silent in the sample is missing from that evidence, not from the genome.

The expected fraction of the rejected hypotheses, or of the accepted calls, that are false. It is not a p value, and controlling it does not make each surviving call certain.

The chance of any false positive among a family of tests. A Bonferroni-style bound suits one gene or one site you intend to assert, not a long list you will follow up.

A record of a header and a sequence. It carries no quality scores and no coordinates, and the free text after the identifier may be false.

A sequence plus a quality string of the same length. The quality characters mean nothing until the encoding, such as Phred+33, is known.

The United States nucleotide archive in that collaboration. A record is not a peer-reviewed fact merely because it was accepted.

A controlled vocabulary with three namespaces, molecular function, biological process, and cellular component. It is not a sequence database, and a term with no evidence code cannot be read.

The DNA of a replicating thing, or an assembly offered as that DNA. A finished genome is still a model, and it does not say what a tissue expresses or what shape a chain adopts.

A display of tracks, gene models, reads, repeats, and variants, on one assembly. Tracks need not agree, and the picture is not a source of truth without the assembly, the track, and the date.

An alignment that scores every residue of both sequences, end gaps included, appropriate when the sequences are homologous along roughly their whole length.

The fast clustering used to decide that order. It is a construction device, not a phylogeny, and publishing it as the family tree is circular.

A genome-wide association study, a test of correlation between genotype and phenotype. A hit is not a cause, and the conventional significance bar for common human variants is a multiple-testing convention, not a biological constant.

A compact, confident predicted fold with nothing in the alignment or the experimental record to justify it, a particular risk when a sequence has no homologues.

A shortcut, such as seed-and-extend search, that is not guaranteed to return the optimal local alignment or every homologue an exhaustive comparison would have ranked as significant.

Proximity data recording which loci were close in the nucleus. Contacts order and orient scaffolds and do not sequence across a repeat; long reads contribute the spanned bases.

Common ancestry. It is not a percentage, so 'forty percent homologous' is a misuse; identity over a stated alignment is evidence from which homology may be inferred, and the inference can be wrong.

A test in which a held-out sequence still has a close homologue in the training set, so the score measures memory. For folds, even a modest identity can carry the answer.

Copying an experimental template after aligning the target to it, then building loops and side chains. The alignment is the model, and energy minimisation does not find a one-residue shift.

A sliding-window average of a hydrophobicity scale, used to propose transmembrane helices. A sustained peak is a candidate, not a proof, and a signal peptide gives a peak that is not a membrane anchor.

The fraction of aligned positions that carry the same letter. The denominator is part of the claim, because counting or ignoring gaps changes the percentage.

Disagreement between a gene tree and a species tree because the ancestor carried more than one allele and speciation happened before those alleles sorted. A high bootstrap then supports a true disagreement, not an estimation failure.

A letter aligned against nothing. The alignment does not say whether one sequence gained the residue or the other lost it.

How far a column departs from equal use of the letters, in bits, with a DNA column that is always one base carrying 2 bits. That quantity describes the column and is separate from a BLAST bit score.

The International Nucleotide Sequence Database Collaboration. A submission to GenBank, ENA, or DDBJ is exchanged with the others, and acceptance checks format rather than biology.

The profile-HMM state that emits extra residues relative to the consensus and can loop, corresponding to extending an insertion.

Removing a sequence or a subtree, realigning it, and keeping the change if the score rises, so an early gap need not be fate. Raising the score cannot rescue an objective that was the wrong one to optimise.

A nucleotide correction that turns an observed fraction of differences into an estimated number of substitutions, allowing for multiple hits. As the observed difference approaches three quarters, the distance goes to infinity.

A string of k consecutive bases or residues. Raising k separates more repeats until errors and shallow coverage break true words.

The statistical model that turns a raw local-alignment score into an E-value, with constants that depend on the scoring system and the background, not on one hit.

The classical hydrophobicity scale used for that plot. A window near nineteen residues is the usual choice for a transmembrane helix, and the plot still has a false-positive rate.

The random-placement model in which the unseen fraction of bases is about e to the power of minus coverage. GC bias and repeats break the assumptions, so the formula is not a promise about an assembly.

Correlation between nearby alleles, so the base that was typed is often a tag for an untyped neighbour. Association then resolves only to the block until an experiment separates the candidates.

An alignment that scores the best-matching segment pair and ignores the rest, appropriate for a shared domain or a read against a long sequence.

A substitution score proportional to the logarithm of how often two residues are paired in trusted alignments, divided by how often that pair would occur by chance. A positive value marks a pair enriched among trusted homologues relative to that background.

A single-molecule read of thousands of bases or more, able to cross a repeat, an isoform, or a junction. In assembly it contributes the bases it spans, not merely a distance, and a systematic error is not removed by averaging.

A tree artefact in which long branches group together because chance matches look like shared changes. Adding sites adds more coincidences and can lock the error in.

Masking of a compositionally biased stretch, such as a glutamine run, before it can seed a hit. A hit that disappears when the filter is on was carried by composition, and its E-value was not a fair test of homology.

In a pairwise alignment, a column in which both sequences have the same letter. In a profile hidden Markov model, the state that emits a residue from a consensus column, as distinct from an insert or a delete.

Preference for the tree and parameter values that give the alignment the highest probability under a stated substitution model. A badly wrong model can make that maximum historically wrong, and more columns then make the error more confident.

A measurement of hybridisation to probes chosen in advance. A transcript with no probe is invisible, and intensity has background and cross-hybridisation.

A column of two different letters, read as a substitution if the homology is accepted.

The hypothesis that substitutions accumulate at a roughly constant rate per unit time across lineages. It is a hypothesis, not a fact about genomes, and it can be rejected.

An arrangement of three or more sequences in which one column proposes positional homology across all of them. It is not the average of the pairwise alignments, which can demand incompatible gaps.

The length of the shortest contig such that contigs at least that long cover half the assembly, after sorting from long to short. It measures contiguity, not correctness, and a false join raises it.

The dynamic-programming algorithm for global alignment. It is not the local algorithm, and it is the wrong default when a short query is compared with a long protein.

A count model whose variance is the mean plus a dispersion term, used because biological replicates of RNA-seq are more scattered than a Poisson model allows.

A distance method that does not assume a clock. Like other distance methods, it does not look back at the alignment once the numbers exist.

A word that does not match a query word exactly but scores above a threshold against it, and is therefore indexed in a protein search.

The classical progressive rule that a gap, once opened, is not reconsidered, so a wrong early order forces later sequences into the same error.

A stretch of codons without a stop. In bacteria a long one is usually a gene; on a eukaryotic chromosome, stop-to-stop is not the gene, because of introns.

A homologue whose split was a speciation. Orthology is not 'the best BLAST hit', and it is not the same event as a duplication.

A lineage added to root a tree. A distant or long-branched outgroup can pull the ingroup, and a root placed at the midpoint assumes roughly equal rates.

Two reads from the two ends of one fragment, recording a distance across an unsequenced middle. In assembly they can name the flanks of a repeat only when the fragment is longer than that repeat.

A family of matrices extrapolated from accepted point mutations counted in closely related proteins. The number, as in PAM250, counts mutational steps, not a percent identity.

A homologue whose split was a gene duplication. Two genes drawn from different species can still be paralogues if the duplication was earlier and each species kept a different copy.

Preference for the tree that explains the columns with the fewest changes. It uses no explicit rate model, which is one way long-branch attraction enters.

The archive of deposited macromolecular coordinate models. A resolution is detail in the map, not a grade for biological truth, and an NMR entry has no crystallographic resolution.

A collection of profile hidden Markov models, one per protein family, each with its own threshold, so a raw score cannot be compared across families.

A quality Q = −10 log₁₀(p), where p is the caller's probability that the base is wrong. It is only as honest as that caller, and it is not a mapping quality.

A per-residue confidence from 0 to 100 on a learned structure prediction. It is not an experimental resolution; below about 50 it should not be read as a stable fold, and a low score does not say whether the cause is disorder or ignorance.

Differences in ancestry between cases and controls that make variants which track ancestry look associated with a disease. A correction cannot rescue a design in which continent and phenotype are the same split.

A table of log-odds for every letter in every column of a multiple alignment, so a near miss at a weak site can still rank high if the rest of the site is exact.

The fraction of positive calls that are true. In the notes' illustration it is about 9% at a prevalence of 1 in 1,000 and about 92% at 1 in 10, with the same sensitivity and specificity.

The tendency of lysine and arginine to lie on the cytoplasmic side of a bacterial inner-membrane protein. It is a hypothesis about orientation, with false calls in both directions.

In a Bayesian tree search, probability proportional to the likelihood times the prior, sampled if the chain has converged. It is a different quantity from a bootstrap percentage, and a tight posterior on a wrong model is not accuracy.

The estimated error in the relative placement of two residues. Two domains can each have a high local confidence while the error between them shows that their packing is not known.

A hidden Markov model of a family, with a match state, an insert state, and a delete state at each consensus column, so gap opening and extension are learned rather than fixed as one penalty.

Building a multiple alignment by aligning the closest sequences first and then aligning alignments, following a guide tree. The order is part of the result.

The proteins in a sample. It is not the translation of the genome, and a database of translated gene models is not a measurement of what the cell contains.

A small prior count added so that a letter never seen in a column is not scored as impossible. Too large a prior erases real conservation.

An iterative protein search that builds a position-specific matrix from one round and searches again. A non-homologue admitted in one round brings its relatives into the next.

The fraction of residues called correctly among three states, helix, strand, and coil. Coil means neither helix nor strand, not disorder, and a string of calls is not a coordinate model.

The screen in which each gene is the other's best match in the other genome. It is not the definition of orthology, and it fails when the true partner is missing, a recent duplicate wins, or several co-orthologues exist.

Under-counting of reads that match the reference less easily, especially around indels, so the pileup can make the reference allele look more common than it is.

The chosen string against which later work is reported. It is a coordinate system, not the ancestor, the healthy state, the common allele, or the organism in front of you.

NCBI's curated layer built from the archive and not identical to it. A predicted-model accession is not the same support for a function as a curated one, and 'in RefSeq' means in that release.

A short motif written as required symbols and alternatives. It matches or it does not, a near miss is rejected, and a hit is not an E-value.

A model that lets the rate change along the tree and still connects change to time. Years enter only through calibrations; sequences do not contain the year.

A count of fragments that map to a gene or a transcript. The library, poly(A) selection or ribosomal depletion, decides which RNAs can appear, and a read in a shared exon does not choose an isoform.

A scaling of fragment counts by transcript length and by sequencing depth. It is not molecules per cell unless an external standard was included, and length does not justify a differential call for the same gene across samples.

A text table of reads placed on a reference, including flag, position, mapping quality, CIGAR, and the sequence. Counting a duplicate or a secondary line invents depth.

A finished chain-termination read of one template, typically several hundred bases, accurate enough that a disagreement is usually biological or a cloning difference rather than a depth problem.

Contigs ordered across gaps, often runs of N whose lengths are estimates from linking data. Quoting those N characters as sequenced bases is a reporting error.

A short word used to find a candidate hit before any full alignment is built. A homologous stretch with no seed above the threshold is invisible to the search.

An alignment that charges internal gaps but not end gaps, fitting a read inside a reference or a shotgun overlap.

A stack of letters in one column of an alignment. The stack height is the information content, and a logo inherits a bad alignment rather than correcting it.

One of a very large number of reads, often about 50–300 bases, whose value is depth. A read shorter than a repeat cannot say which copy it came from.

Sequencing of random fragments, with order inferred from overlaps rather than from a directed walk along the chromosome.

A cleaved hydrophobic stretch at the N-terminus of a secreted protein. Reading it as a transmembrane helix invents a membrane anchor.

A loose word for identical letters, conservative substitutions, or sometimes a raw score. It is not a standard percentage and not a probability, and it should be replaced by a named score, an identity count, and a coverage.

A multiplier that removes the effect of sequencing depth, taken from the typical gene rather than from a total that a few dominant transcripts can capture.

The dynamic-programming algorithm for local alignment. It does not let the score stay negative, and traceback starts at the best cell rather than at the corner.

A variant large beside a short read, seen only indirectly by pairs at the wrong distance, by split reads, or by a change in depth. Its absence from a file of single-base calls is not evidence it is absent from the genome.

A coordinate set fitted to an experiment under stated conditions, or a prediction that shares a file format and not the evidence. The shared format does not make them the same kind of result.

The reviewed part of the protein archive, in which a curator has compared the sequence with the literature. The review can be incomplete, and it is not itself a new assay.

The experimental structure whose backbone is copied. No named template means there is nothing of this kind to build, however low an energy a program reports.

A length-adjusted abundance rescaled so that the values in one sample sum to one million. It is a proportion within the library, not a count of molecules per cell.

The walk back through the choices that won each cell, recovering the alignment itself rather than only its score. A tie means more than one alignment is optimal.

The RNA of one sample, a cell type, a time, or a treatment. There is no transcriptome of a species in the way there is, approximately, a genome, and an RNA abundance is not a protein abundance.

The copying of a function from a label that was itself copied, so that a name such as kinase can spread with no assay anywhere in the chain.

The unreviewed protein entries, automatic translations whose product names are usually transferred from other names. The accession is stable; the mnemonic name can change.

The low-identity range, in these notes below about 25–30% over a long globular alignment, where homology may be real and an automatic model is not evidence. The bands are a rule of thumb, not a law.

The tree that sequences alone can support, with no direction of time. A named clade is a claim about a rooted tree.

A distance method that forces equal root-to-tip lengths, which is a molecular clock, and that should not be used when rates differ.

A line naming a variant call against a stated reference. A filter value of PASS means this caller's filters accepted the site, not that the variant is real or pathogenic.

The revision that fixes one sequence under an accession, and, separately, the program release that must be named if a search is to be repeated. Repeating yesterday's commands on this morning's database is a new experiment.