Glossary

Terms as CRAIC uses them. Where a word is used differently elsewhere in the literature, that is noted, because most of the confusion around alignment reliability is vocabulary rather than statistics.

Alignment — a rectangular arrangement of sequences in which each column asserts that its residues descend from the same ancestral position. Every column is a hypothesis; see homology.

Ambiguous region — a stretch where the evidence does not strongly prefer one arrangement over its neighbours. Usually created by insertions and deletions: substitutions alone rarely make a column doubtful.

Codon-aware alignment — aligning protein-coding nucleotides by translating, aligning the amino acids, and threading the nucleotides back through the result, so gaps land on codon boundaries and the reading frame survives.

Combined score — CRAIC's headline per-column reliability: the mean of the consistency and perturbation scores. In [0, 1].

Conservation — how similar the residues in a column are. Not reliability: a column can be perfectly conserved and completely misplaced. CRAIC shows it as a separate track for exactly this reason.

Consistency score — per column, the mean pair-HMM posterior over the residue pairs the column groups together. High means the unaligned-sequence evidence supports the column. In the spirit of T-Coffee's TCS and of ZORRO, but computed over CRAIC's own posteriors and so not numerically interchangeable with either.

Consistency transformation — the ProbCons idea of re-estimating each pairwise posterior using every third sequence, so that evidence from the whole family informs each pair. Costs O(N³) in sequences and buys a large accuracy improvement.

Core block — in a structural reference set such as BAliBASE, the region the reference is actually confident about. Published BAliBASE SP and TC figures are restricted to core blocks, so an all-column figure is not comparable with them.

Correct placement — CRAIC's per-residue view of accuracy against a known answer: the fraction of a residue's partners in its column that are genuinely homologous to it. The per-residue counterpart of reference correctness, and exactly consistent with it — a column's score is the mean of its residues'.

Disagreement / aligner agreement — how often the other installed aligners reproduce each column of the current alignment. A model-free proxy for uncertainty: where methods differ, something is being guessed.

Gappyout / strict — trimAl's automatic column-selection rules, reimplemented in craic.ambiguity.trimming. They choose their own threshold from the distribution of gap scores rather than taking one from the user.

Genetic code — the NCBI translation table used to turn codons into amino acids. CRAIC carries one per alignment and defaults to table 1 (standard). Mitochondrial and Mollicute genomes differ, most consequentially in reading TGA as tryptophan rather than as a stop.

Guide tree — the order in which a progressive aligner merges sequences. Errors in it are a dominant source of alignment error, which is why perturbing it is informative.

Homology — here, positional homology: two residues are homologous if they descend from the same ancestral position. A column asserts homology among all its residues; SP and TC both count homologous pairs and columns respectively.

Internal stop codon — a stop before a sequence's last codon. Usually means the wrong genetic code, the wrong reading frame, or a pseudogene; CRAIC reports them but cannot say which.

Intruder — in the column inspector, a residue placed in a column that the reference puts somewhere else.

Masking — removing columns judged unreliable before a downstream analysis, or, with a residue mask, only the residues judged unreliable (written as missing data, N or X, with the column kept). CRAIC previews it live (dimmed columns) and exports the result. Whether it helps is an empirical question, and a threshold that removes hard columns will raise mean accuracy whatever rule picks them — hence the random and gap-matched controls in the benchmark.

MEA (maximum expected accuracy) — decoding an alignment that maximises the expected number of correctly aligned residue pairs under the posterior, rather than one that maximises a sum of substitution scores minus gap penalties. Gap placement falls out of the model instead of being tuned.

Missing residue — in the column inspector, a residue the reference puts in the column being inspected but which the alignment placed elsewhere. A column can score a perfect 1.0 and still have one: the score asks whether the homologies a column asserts are true, and a column that left a residue out asserts nothing false.

Pair-HMM — the three-state (match / insert / delete) hidden Markov model with affine gaps whose forward–backward posteriors are the single primitive everything in CRAIC is built on.

Perturbation score — per column, the fraction of its asserted homologies that survive re-alignment across an ensemble of bootstrapped guide trees and gap regimes. Inspired by GUIDANCE but not equivalent to it: 16 replicates of CRAIC's own engine, not ~100 of the aligner that built the alignment.

Posterior (match probability) — P[i, j], the probability that residue i of one sequence is homologous to residue j of another, under the pair-HMM. The posterior explorer draws this matrix directly.

Precision — of an alignment against a reference, the fraction of the homologous pairs it asserts that are true. The complement of SP, which is recall. Reported because a gappy alignment can have high precision and terrible recall.

Reference alignment — a trusted alignment of the same sequences: a structural reference such as BAliBASE, or the known-true alignment of a simulated dataset.

Reference correctness — per column, the fraction of the homologous pairs it asserts that are true, given a reference. nan for a column that asserts no pair at all, which can be neither right nor wrong.

Reliability — how well supported a column is, as distinct from how conserved it is. CRAIC's reliability is the combined score.

Reliability AUC — the ROC area under the curve of a reliability score predicting which columns are actually correct, given a known answer. 0.5 is chance. This is the number that says whether the reliability claim is worth anything on a given dataset.

Session — CRAIC's own document (*.craic.json): the alignment plus the reference, annotations, sequence groups, provenance log and view state. A FASTA export keeps the residues and drops all of that.

SP (sum-of-pairs) score — the fraction of the reference's homologous residue pairs that the alignment recovers. Recall, in other words.

TC (total column) score — the fraction of the reference's columns that the alignment reproduces exactly. A much harsher measure than SP: one misplaced residue loses the whole column.

Truth mode — working with a reference attached, so that accuracy is scored and displayed live. CRAIC's teaching centre of gravity.