Glossary: Ensembl, Galaxy, AI, Prompt Engineering & RAG

A reference glossary spanning genome resources, workflow platforms, and the generative-AI concepts used to build training and analysis tools around them.


1. Ensembl

Ensembl — A genome browser and annotation database jointly run by EMBL-EBI and the Wellcome Sanger Institute, providing gene sets, variation, comparative genomics and regulatory data across vertebrate genomes.

Ensembl Genomes — A sister project extending the Ensembl platform to non-vertebrate taxa, divided into Ensembl Plants, Fungi, Metazoa, Protists and Bacteria. Ensembl Plants hosts Oryza sativa and other crop genomes.

Assembly — A specific reconstructed version of a genome sequence (e.g. GRCh38, IRGSP-1.0 for rice). Annotations are tied to an assembly version.

Gene annotation / gene set — The catalogue of genes, transcripts and their structures predicted or curated for an assembly, released as the “Ensembl gene set.”

Stable ID — A persistent identifier for a feature: ENSG (gene), ENST (transcript), ENSP (protein). Species-specific prefixes exist (e.g. OS for rice in Ensembl Plants).

BioMart — A data-mining interface for bulk querying and exporting Ensembl data by filters and attributes without writing code.

VEP (Variant Effect Predictor) — A tool that annotates variants with their predicted consequences on genes, transcripts and proteins (e.g. missense, stop-gained), plus regulatory and population-frequency context.

Compara — The comparative-genomics database providing gene trees, homology relationships and whole-genome alignments across species.

Homologue — A gene related to another by common ancestry; split into orthologues (related by speciation, across species) and paralogues (related by duplication, often within a species).

Gene tree — A phylogenetic tree representing the evolutionary history of a gene family, used to infer orthology and paralogy.

Regulatory Build — Ensembl’s genome-wide annotation of regulatory features (promoters, enhancers, transcription-factor binding sites) derived from epigenomic data.

Variation database — The component storing known sequence variants (SNVs, indels, structural variants) and their population frequencies.

GTF / GFF — Tab-delimited text formats describing gene and feature coordinates; the standard way Ensembl annotation is exported for pipelines.

REST API — Ensembl’s programmatic interface for retrieving sequences, features and annotations over HTTP, suitable for scripting and automation.


2. Galaxy Workbench

Galaxy — An open-source, web-based platform for accessible, reproducible data-intensive bioinformatics, letting users run tools and pipelines without command-line access.

Galaxy instance — A deployed Galaxy server. Public examples include usegalaxy.org (US), usegalaxy.eu (Europe) and usegalaxy.org.au; institutions also run private instances.

History — A user’s working record of datasets and the operations applied to them, forming an auditable analysis session.

Dataset — Any input, intermediate or output file tracked within a Galaxy history, with associated metadata and provenance.

Dataset collection — A structured group of datasets (e.g. a list of samples, or paired reads) processed together as a single logical unit.

Tool — A wrapped command-line program exposed in Galaxy through a graphical form for setting parameters.

Tool wrapper — The XML definition that maps a command-line tool’s inputs, parameters and outputs into Galaxy’s interface.

Tool Shed — Galaxy’s app store for discovering and installing community- contributed tools and dependencies into an instance.

Workflow — A saved, reusable chain of tools that connects outputs to inputs, enabling automated, repeatable multi-step analyses.

Invocation — A single execution run of a workflow, recording the inputs, parameters and resulting datasets.

Job — The unit of computation Galaxy dispatches to a backend (local, cluster or cloud) when a tool runs.

Interactive tools — Live environments such as Jupyter or RStudio launched inside Galaxy for exploratory analysis alongside standard tools.

Provenance / reproducibility — Galaxy’s automatic capture of every tool, version, parameter and dataset so an analysis can be rerun or shared exactly.

Conda / BioConda integration — The dependency-resolution system Galaxy uses to install reproducible tool environments.

Galaxy Training Network (GTN) — A community-maintained collection of hands-on tutorials and training materials built on Galaxy.


3. Genome Annotation & Sequence Characterization

Structural annotation

Genome annotation — The process of identifying functional elements in a genome and describing what and where they are. It splits into structural annotation (locating features) and functional annotation (assigning roles).

Structural annotation — Determining the location and architecture of features such as genes, exons, introns and untranslated regions along the sequence.

Gene prediction (gene calling) — Computationally identifying gene locations and structures, either ab initio (from sequence signals alone) or evidence-based (guided by RNA-seq, ESTs or protein alignments).

Ab initio prediction — Gene prediction from intrinsic sequence features (codon usage, splice signals) using trained models such as AUGUSTUS or GeneMark.

Evidence-based / homology-based annotation — Predicting gene models by aligning external evidence — transcripts, proteins from related species — to the genome.

Annotation pipeline — An integrated workflow (e.g. MAKER, BRAKER) that combines repeat masking, evidence alignment and gene predictors into consensus gene models.

Gene model — The defined structure of a gene: its exons, introns, coding sequence and untranslated regions, including alternative transcripts.

ORF (open reading frame) — A stretch of sequence between a start and stop codon with the potential to encode a protein.

CDS (coding sequence) — The portion of a transcript that is translated into protein, excluding untranslated regions.

Exon / Intron — Exons are sequence segments retained in the mature transcript; introns are removed by splicing.

UTR (untranslated region) — Transcribed but non-coding regions flanking the CDS (5′ and 3′ UTRs) that influence regulation and stability.

Splice site — The sequence boundary (canonically GT…AG) marking where introns are removed and exons joined.

Repeat masking — Identifying and flagging repetitive sequence (e.g. with RepeatMasker) before gene prediction; soft masking lowercases repeats while hard masking replaces them with Ns.

Transposable element — A mobile, often repetitive DNA sequence; abundant in plant genomes and a major focus of repeat annotation.

ncRNA annotation — Identifying non-coding RNAs such as rRNA, tRNA, miRNA and lncRNA, often with specialised tools (e.g. tRNAscan-SE, Infernal/Rfam).

Pseudogene — A sequence resembling a functional gene but disabled by mutations; flagged during annotation to avoid false gene calls.

Annotation transfer / liftover — Projecting annotations from one assembly or species onto another via alignment or synteny, useful for newly assembled genomes.

Synteny — Conservation of gene order and content across genomes, used to support annotation transfer and comparative analysis.

GFF3 — The standard hierarchical text format for representing structural annotation (gene → mRNA → exon/CDS) with coordinates and attributes.

Functional annotation & sequence characterization

Functional annotation — Assigning biological meaning to features — molecular function, biological process, pathways — typically by transferring information from characterised homologues.

Homology search — Finding similar sequences in databases (e.g. BLAST, DIAMOND, MMseqs2) to infer function or evolutionary relationship.

Sequence alignment — Arranging sequences to maximise correspondence; pairwise compares two, multiple sequence alignment (MSA) compares many to reveal conserved positions.

Profile HMM (Hidden Markov Model) — A statistical model of a sequence family capturing position-specific conservation, used to detect distant homologs and domains.

Protein domain — A distinct, often independently folding functional unit within a protein; conserved domains are strong functional clues.

Motif — A short, recurring sequence pattern associated with a function, such as a binding site or catalytic site.

InterPro / InterProScan — A resource (and its scanning tool) that integrates multiple signature databases (Pfam, PROSITE, etc.) to assign domains and functional sites.

Pfam — A widely used database of protein families represented as profile HMMs.

Gene Ontology (GO) — A controlled vocabulary describing gene products across three aspects: molecular function, biological process and cellular component.

KEGG / pathway annotation — Mapping genes to metabolic and signalling pathways using resources such as KEGG.

EC number — A four-part Enzyme Commission identifier classifying an enzyme by the reaction it catalyses.

Signal peptide / transmembrane prediction — Identifying localisation and membrane-spanning features from sequence (e.g. SignalP, TMHMM/DeepTMHMM), relevant to characterising proteins like ER-localised receptors.

BUSCO — A completeness assessment that checks for expected single-copy orthologues to gauge how complete an assembly or annotation is.

GC content — The proportion of guanine and cytosine bases in a sequence; a basic compositional descriptor used in characterization and quality checks.

k-mer — A substring of length k; k-mer frequencies underpin many assembly, classification and characterization methods.

Codon usage — The relative frequency of synonymous codons in coding sequence, used in gene prediction and expression analysis.

Annotation evidence / confidence — Metadata indicating how a feature was supported (experimental, computational, homology), important for judging reliability of a gene model.

4. Artificial Intelligence (Core ML Concepts)

Artificial Intelligence (AI) — The broad field of building systems that perform tasks normally requiring human intelligence, such as reasoning, perception or language understanding.

Machine learning (ML) — A subset of AI where systems learn patterns from data rather than being explicitly programmed with rules.

Deep learning — ML using multi-layered neural networks to learn hierarchical representations directly from raw data.

Neural network — A model of interconnected nodes (“neurons”) organised in layers, whose connection strengths (weights) are adjusted during training.

Parameters / weights — The learned numerical values inside a model that encode what it has learned; modern LLMs have billions of them.

Training / pretraining — The process of fitting model parameters to data. Pretraining builds general capability on large corpora before any task- specific adaptation.

Fine-tuning — Further training of a pretrained model on a narrower dataset to specialise it for a task or domain (e.g. adapting a genomics model to rice).

Inference — Using a trained model to generate predictions or outputs on new inputs, as opposed to training it.

Supervised vs. unsupervised learning — Supervised learning uses labelled examples; unsupervised learning finds structure in unlabelled data.

Large language model (LLM) — A neural network trained on vast text to predict and generate language, capable of summarising, reasoning and answering questions.

Transformer — The neural-network architecture underlying modern LLMs, built around an attention mechanism that weighs the relevance of different parts of the input.

Foundation model — A large, general-purpose model pretrained on broad data that can be adapted to many downstream tasks.

Token — The unit of text a model processes (a word, sub-word or character fragment); models read and generate text one token at a time.

Embedding — A numerical vector representing the meaning of a token, sentence or document, where similar items sit close together in vector space.

Context window — The maximum amount of text (in tokens) a model can consider at once, including both the prompt and its response.

Hallucination — A confident but factually incorrect or fabricated model output, a key risk to mitigate in scientific applications.

Multimodal — A model able to process more than one data type, e.g. text and images together.

Agent / agentic AI — An AI system that plans and takes multi-step actions toward a goal, often calling external tools, with optional human-in-the-loop oversight.


5. Prompt Engineering

Prompt — The input text given to a model that frames the task and provides any instructions, context or examples.

Prompt engineering — The practice of designing and refining prompts to elicit accurate, reliable and well-formatted model outputs.

System prompt — A high-level instruction that sets a model’s role, constraints and behaviour for an entire session, separate from user messages.

Zero-shot prompting — Asking a model to perform a task with no worked examples, relying solely on instructions.

Few-shot prompting — Providing a handful of example input–output pairs in the prompt to demonstrate the desired behaviour.

In-context learning — A model’s ability to adapt to a task purely from examples and instructions in the prompt, without updating its weights.

Chain-of-thought — Prompting a model to reason step by step before answering, which improves performance on multi-step problems.

Role / persona prompting — Instructing the model to adopt a specific viewpoint or expertise (e.g. “act as a genomics curator”) to shape its tone and focus.

Prompt template — A reusable, parameterised prompt structure with slots for variable inputs, enabling consistent automated use.

Structured output — Constraining a model to return a defined format such as JSON or a table, so its output can be parsed programmatically.

Temperature — A setting controlling output randomness; lower values give deterministic, focused responses, higher values give more varied ones.

Prompt chaining — Breaking a complex task into a sequence of prompts where each step’s output feeds the next.

Guardrails — Rules, filters or constraints applied to inputs and outputs to keep model behaviour safe, on-topic and reliable.

Prompt injection — A security attack where malicious instructions hidden in user input or external content try to override the intended prompt; a key concern when models read untrusted data.


6. Retrieval-Augmented Generation (RAG)

Retrieval-augmented generation (RAG) — An architecture that retrieves relevant documents from an external knowledge source and supplies them to an LLM as context, grounding its answers in specific, up-to-date material.

Knowledge base — The curated collection of documents (papers, protocols, annotations) that a RAG system retrieves from.

Chunking — Splitting source documents into smaller passages so they can be embedded, indexed and retrieved at a useful granularity.

Embedding model — A model that converts text chunks and queries into vectors for similarity comparison; distinct from the generation LLM.

Vector database — A store optimised for indexing embeddings and performing fast nearest-neighbour similarity search (e.g. FAISS, Chroma, Pinecone, Weaviate).

Indexing — Preprocessing the knowledge base into embeddings and data structures so retrieval is efficient at query time.

Semantic search — Retrieval based on meaning via embedding similarity, rather than exact keyword matching.

Top-k retrieval — Returning the k most relevant chunks for a query to pass into the model’s context.

Hybrid search — Combining dense (embedding-based) retrieval with sparse keyword methods such as BM25 to improve recall and precision.

Re-ranking — A second-pass step that reorders retrieved candidates by relevance, often with a more accurate but costlier model.

Grounding — Tying a model’s response to retrieved evidence so claims are traceable to source documents, reducing hallucination.

Context augmentation — Injecting retrieved passages into the prompt so the model generates an answer informed by them.

Citation / attribution — Linking parts of a generated answer back to the specific source passages that support them, important for scientific trust.