Bioinformatics and AI Glossary

Sequencing and data

Read — a single sequence output by a sequencer.
Coverage / depth — average number of reads spanning each base.
Phred score (Q) — log-scaled base error probability; Q30 = 1 in 1000.
Contig / scaffold — contiguous assembled sequence / ordered contigs with gaps.
N50 — length such that half the assembly lies in contigs of at least that size.
FASTQ / BAM / VCF / GFF — reads / alignments / variants / annotations (Sequencing Data Formats).

ANI — average nucleotide identity; ~95% ≈ species boundary.
Core / accessory genome — genes in nearly all vs some strains (Pangenome Analysis).
Ortholog / paralog — same gene in different species / duplicated within a genome.
MGE — mobile genetic element (plasmid, transposon, IS, prophage, integron).
Replicon / Inc type — self-replicating DNA unit / plasmid incompatibility group.
ST / cgMLST — sequence type / core-genome allele-based type (MLST and cgMLST).
tMRCA — time to most recent common ancestor (Phylodynamics).
MAG — metagenome-assembled genome (Metagenome-Assembled Genomes).
ASV / OTU — amplicon sequence variant / operational taxonomic unit (16S Amplicon Analysis).

Supervised / unsupervised — learning with / without labels.
Feature — one measured input variable; embedding — learned dense representation.
Overfitting — memorizing training data; fails on new data.
Data leakage — test information contaminating training; the top cause of inflated results.
Cross-validation — repeated internal splits; weaker than external validation.
AUC-ROC / AUC-PR — ranking performance; PR is preferred under class imbalance.
Calibration — agreement between predicted probability and observed frequency.
Sensitivity / specificity / PPV / NPV — see Model Evaluation in Clinical Microbiology.
Very major error — genotype/model says susceptible, phenotype is resistant (most dangerous).
SHAP — per-prediction feature attribution for interpretability.
Drift — performance decay as populations or protocols change.

Neural network / layer / weights — stacked learned transformations.
CNN — convolutional network, for images.
Transformer / attention — architecture behind protein and language models.
GNN — graph neural network, for molecules and graphs.
Pretraining / fine-tuning — general learning then task adaptation.
Zero-shot — usable without task-specific labels (Protein Language Models).
pLDDT / PAE — AlphaFold confidence measures (AlphaFold in Microbiology).
Hallucination — confident but fabricated model output (Foundation Models and LLMs in Microbiology).
RAG — retrieval-augmented generation; grounding answers in retrieved sources.

Container — packaged software environment (Docker/Singularity).
Workflow manager — Nextflow/Snakemake; reproducible multi-step pipelines.
FAIR — findable, accessible, interoperable, reusable (FAIR Data and Genomic Surveillance).
Provenance — recorded tool/database versions and parameters behind a result.