Glossary

Terms are grouped by topic and listed alphabetically within each group.

Genomics

Accession (ENCFF, ENCSR). ENCODE's stable identifiers. ENCFF... names a file and ENCSR... an experiment. For example, ENCFF381NDD is a K562 H3K27ac fold-change-over-control bigWig from experiment ENCSR000AKP.

Assembly (genome build). A specific version of a reference genome sequence, which fixes the coordinates. hg38 (UCSC name) and GRCh38 (Genome Reference Consortium name) are the same human assembly; hg19 is the older GRCh37. In Genie the genome is fixed per project, and hg38 is sent to ENCODE as GRCh38. See Genome assemblies.

ATAC-seq. An assay that maps accessible chromatin by inserting sequencing adapters into open DNA.

BED. A plain-text format for genomic intervals: chromosome, start, end, then optional fields such as name and score. Intervals are 0-based and half-open, so chr8 100 110 covers 10 bases. Genie's peak tool writes its results as BED files.

bedGraph. A plain-text format that gives a numeric value for each interval (chromosome, start, end, value). Genie treats bedGraph tracks as signal tracks.

bigBed. The indexed binary form of BED, for large interval sets.

bigWig. An indexed binary format for continuous signal along the genome. Only the parts of the file needed for a region are read, so a browser can display remote files efficiently. Most ENCODE signal tracks Genie loads are bigWigs.

cCRE (candidate cis-regulatory element). An entry in the ENCODE Registry of candidate regulatory regions, defined from DNase accessibility and histone-mark data.

Cell lines (K562, HepG2). Populations of cells that grow indefinitely in culture and are used as shared model systems. K562 is a leukemia cell line, derived from a patient with chronic myelogenous leukemia. HepG2 is a liver cancer cell line; Cellosaurus records it as hepatoblastoma-derived (CVCL_0027). In the ALB demonstration, "liver" and "blood" stand only for these two cell lines.

ChIP-seq. Chromatin immunoprecipitation followed by sequencing. An antibody pulls down DNA bound to a protein or carrying a histone mark; the DNA is sequenced and mapped to the genome. Peaks are usually called with model-based tools such as MACS.

Chromatin accessibility. How open a stretch of DNA is to enzymes and proteins. Accessible regions often coincide with promoters and enhancers. Measured by DNase-seq or ATAC-seq.

Control (input). A matched sample processed without the specific antibody, used to correct ChIP-seq signal for background.

DNase-seq. An assay that maps DNase I hypersensitive sites, which are accessible chromatin. ENCODE's DNase-seq signal files are "read-depth normalized signal"; ENCODE has no DNase-seq files of output type "fold change over control".

ENCODE. The Encyclopedia of DNA Elements, a consortium and data portal of functional genomics experiments. It is the only portal Genie's search tool queries. See Public data search.

Enhancer. A DNA region that increases transcription of a gene, often from a distance and in a cell-type-specific way. Which regions count as enhancers depends on the evidence used (chromatin marks, accessibility, or functional tests such as CRISPR interference), so say which evidence supports a call.

Fold change over control. An ENCODE signal output type: at each position, the ChIP signal divided by a background estimated from the input control. A value of 1 means the same as control. Genie's search asks for this output type first.

H3K27ac. Acetylation of lysine 27 on histone H3. It marks active enhancers and active promoters. H3K27ac alone does not tell a promoter from an enhancer; position relative to a transcription start site, or other marks, is needed.

H3K4me3. Tri-methylation of lysine 4 on histone H3, enriched at active promoters.

Histone modification (histone mark). A chemical tag on the histone proteins that package DNA. Names follow the pattern histone, residue, modification: H3K27ac is histone H3, lysine (K) 27, acetylation.

Locus. A position or region of the genome, often named after a gene ("the MYC locus"). Genie writes a locus as chr:start-end, for example chr8:127,700,000-127,760,000.

Locus control region (LCR). A group of regulatory elements that drives high-level expression of a gene cluster. The beta-globin LCR is the classic example.

Peak. A region where signal is higher than background. What counts as a peak depends on the method: model-based callers use the control and report significance, while Genie's peak tool is a simple threshold caller.

Promoter. The DNA region at and just upstream of a gene's transcription start site, where transcription begins. Exact boundaries are a convention; state the window used when calling something a promoter.

Read-depth normalized signal. An ENCODE output type in which signal is scaled by the number of sequenced reads but is not a ratio to a control.

Track hub. A remote collection of tracks described by small configuration files that a browser can load by URL. Genie's agent cannot load a hub URL.

Transcription start site (TSS). The first base transcribed into RNA. A gene can have several TSSs, one per isoform, so say which annotation (for example GENCODE or MANE) was used.

Analysis and statistics

Bin, binning. Dividing a region into equal-width windows and summarizing the signal in each. Genie's correlation tool uses about 500 bins for regions of 50 kb or more; empty bins count as 0.

Fragment (peak fragment). A peak interval that is part of a larger enriched region, split off because Genie's peak caller merges only overlapping or touching intervals. Genie's peak counts are fragment counts, not counts of regulatory elements.

Mean plus 2 SD. Genie's default peak cutoff: the mean of the fetched interval scores plus 2 sample standard deviations. The peak tool's output field threshold holds this absolute cutoff, not a z value.

Pearson correlation (r). A number from -1 to 1 that measures how well two sets of values fit a straight line. Computed over bins along the genome, it depends on the bin size, and neighbouring bins are not independent, so a standard significance test overstates the evidence. Multiplying a track by a positive constant does not change r. One window's r measures co-variation across bins, not agreement between assays or tissue specificity.

Scope. The region a Genie computation tool works over: viewport (the default), chromosome or all, and for signal statistics also explicit regions. The four computation tools refuse scopes over 5 Mb. See Measurements.

Threshold peak calls. Genie's term for the output of its peak tool: intervals at or above a cutoff, with no control, p-value or false discovery rate. Treat them as exploratory.

Genie terms

Agent mode. The type of an agent chat, with its own tools and instructions: Genome mode, Analysis mode and Asset import mode. A chat is offered the Genie tools of one mode only. See Modes.

Asset. A file or remote source registered to a project, of one of 7 kinds. Importing an asset does not add a browser track. See Assets.

Chip (citation chip). The numbered marker ([1], [2], ...) the chat shows in place of a citation token it can resolve. Numbering restarts in each message. Clicking a chip opens the evidence panel.

Citation token. A string of the form [[cite:<id>]] that a Genie tool returns with its result. The agent is told to copy it after the claim it supports and never to invent one. The tools, not the model, issue citation tokens. A token whose id is not in the chat's records stays as raw text. See Citations and evidence.

Declared tool. A tool Genie defines (name, description, input schema) and passes to the agent runtime when a chat starts. Genome mode declares 14, Analysis mode 6 and Asset import mode 8, 23 distinct. The runtime's own built-in tools are not declared tools. See Tool reference.

Discovery Studio. The Analysis-mode area that lists saved Studio outputs.

Evidence panel. The dialog that opens when you click a chip. It shows the stored evidence record: its type, title, location, token, genome and excerpt (for computation tools, the tool's result JSON up to 1,800 characters), with an Open source link where a URL exists.

Evidence record (stored evidence). What a citation token points to: a computation tool's result, an ENCODE search or file record, an asset record, or, for navigation and the viewport summary, only the region.

Project. A workspace with a fixed genome, its assets, browser state and chat sessions. See Projects and genomes.

Report. A Studio output of kind "report", used for a saved analysis plan. See Plans and outputs.

Session (chat). One chat thread with the agent in one mode, with its own folder holding its citation records (citations.json) and runtime log (app-server.log).

Studio output. An HTML document the agent saves in Analysis mode, of one of 10 kinds, such as a report, data table or hypothesis cards.

Viewport. The genomic region currently shown in the browser. Many Genie tools default to the viewport.

Computing terms

Agent runtime. The software that runs the model loop: it sends the conversation to the model, passes tool calls back to the app and returns their results. Genie uses the OpenAI Codex CLI app-server.

App-server. The Codex CLI's JSON-RPC server mode, started as codex app-server. Genie's backend starts a thread for each chat (declaring tools, instructions and the model) and sends each user message as a turn.

Approval policy, auto-accept. Whether the runtime must ask before certain actions, such as shell commands. Genie's backend is configured to auto-accept the runtime's approval requests. See Runtime and safety.

Built-in runtime tool. A capability supplied by the agent runtime itself rather than by Genie, such as shell commands or web search. Genie does not switch these off.

eg3. The engine of the third-generation WashU Epigenome Browser, which renders Genie's browser panel. It is free for non-commercial use under its own licence. See Browser engine.

JSON-RPC. A simple protocol for remote procedure calls encoded as JSON. Genie's backend talks to the Codex CLI app-server with JSON-RPC over standard input and output.

LLM (large language model). A neural network trained on large text corpora, which can follow instructions and hold a conversation. An LLM agent runs the model in a loop in which it can call tools, read their results and decide the next step.

MCP (Model Context Protocol). An open protocol for connecting LLM applications to external tools and data. Genie does not implement an MCP server; it passes its tools to the runtime directly. The runtime can separately start MCP servers from the user's own configuration.

Reasoning effort. A runtime setting for how much the model reasons before answering (for example low or medium). You choose it in the chat composer.

Sandbox. Limits the runtime places on what the agent's commands can do. Genie's per-turn policy is workspace-write with network access on, with the chat's folder as the writable root.

Tool call. One request by the model to run a named tool with given arguments, and the result the tool returns. The chat shows each call as a row with its status.

WebSocket. A persistent two-way connection between the browser and a server. The Genie interface talks to the local backend over a WebSocket at /ws (default port 8787).